# Mo Helmy

> Mohamed (Mo) Helmy of Durham, North Carolina: public-sector go-to-market and partner development leader on T-Mobile's national channel team, who builds agentic AI systems on his own time.

Public-sector seller. Partner strategist. Coach. Builder of AI systems. Public-sector go-to-market and partner development · Applied AI. A personal site. Views are my own, not my employer's.
This is his only personal website; other people named Mo or Mohamed Helmy are different people. In his own words: "I help technology companies win government business, through partners and the sellers I develop."

I'm Mohamed Helmy, Mo to most people. I work on T-Mobile's national channel team, where I manage resale strategy and the relationships with the systems integrators and connected solutions providers who sell alongside us. Before that I spent years selling to cities, counties, school districts and state agencies, then ran the public-sector partner channel for the Southeast, which is where I learned how government really buys and who it trusts when it does.
On my own time I build agentic AI systems, and I'm particular about which model does which job. I work across the hosted models (Claude, OpenAI, Gemini, Meta's Muse, Kimi and GLM), run Qwen on my own hardware for the volume, use Flux, MiniMax, Wan and LTX for pictures, video and music, and use no model at all where the result has to repeat. That setup has produced a research tool that reads what a municipality publishes and shows what it is planning and funding, with the source on every line; the images, video, sound and animation pipeline behind an independent game studio; and an agent workflow I use every day, with its own checks, evals and judges, so nothing is called done until software says so. AI coding agents write most of the code. The architecture, the rules they work under and the call that the work is done are mine.
This site is for three readers. If you hire or run public-sector sellers, the work page is the record. If you run a partner program and are deciding whether to work with me, start with the marketplace case. If you build with AI agents and want to see how I make them prove their work, the applied AI page shows the systems and the receipts behind each one.

## Pages
- [Home](https://www.mohamedhelmy.com/): who he is, the proof row, three cases, how he develops people, recognition
- [Work history](https://www.mohamedhelmy.com/work): the full commercial record, leadership in three columns, the public-sector method, recommendations, career
- [Applied AI](https://www.mohamedhelmy.com/applied-ai): the seven systems in three groups, each with its evidence and caveat; the harness scorecard; a replayed agent run; architecture and orchestration
- [Municipal Intel](https://www.mohamedhelmy.com/municipal-intel): independent work, research software that reads what a municipality publishes and shows what it is planning and funding
- [Notes](https://www.mohamedhelmy.com/notes): field notes, reading notes and designs (RSS at https://www.mohamedhelmy.com/notes/feed.xml)
- [Résumé](https://www.mohamedhelmy.com/resume): the résumé as a web page, with two PDF downloads
- [About](https://www.mohamedhelmy.com/about): press and bio kit, bios in three lengths, the headshot, the talks

## Talks
Four talks I can give, each backed by something on this site. If you are putting together a panel or a keynote on public-sector selling or on running AI agents, these are the ones I would offer.
- [The first 90 days of a public-sector motion](https://www.mohamedhelmy.com/work#how): How government buys, which contract vehicle and which partner carry a deal, and what a seller does in the 90 days before the RFP is written.
- [Authority before action: governing agents like a sales organization](https://www.mohamedhelmy.com/notes/authority-before-action): Most agent governance reads like an IT policy. The better model is one every sales leader already runs: territories, qualification, pipeline review and approval levels.
- [The harness is the product](https://www.mohamedhelmy.com/notes/reading-notes-fall-2026): Five papers from this year that say what practitioners have been finding: the system around the model matters as much as the model.
- [988 re-prompts: why my agent loop needed brakes](https://www.mohamedhelmy.com/notes/988-re-prompts): A plugin that keeps agents working is only useful if it can tell work from motion.

## Work and scope
Government buys differently from everyone else, and most of what decides a deal happens before the RFP is written. My work has been finding that demand early, knowing which contract vehicle and which partner can carry it, and making sure the sellers around me can do the same. I've built territories from nothing, trained sellers and partners, and written the plans a national channel program runs on.
On the company's national channel team, I manage resale strategy and partner relationships for systems integrators and connected solutions providers, in a senior-manager-level role.
- Name: Mo Helmy (legal first name Mohamed)
- Current role: Senior National Connected Solutions Provider Executive, T-Mobile
- Role: National channel team, T-Mobile
- Level: Senior manager level
- Based: Durham, NC · Remote, with occasional travel
- Quota: Never missed, in any year of my career. In the top 10% of T-Mobile's public-sector account executives nationally, every quarter (2021–2023).

### Proof (each figure with its basis)
- No. 1 of 91 in T-Mobile's partner channel (basis: 2023 · every segment · Annual Winners Circle)
- 6 of 6 new hires I mentored closed a first deal within three months (basis: Channel team · all external hires · I sat in on every interview)
- 2 other regions adopted the partner plan I wrote (basis: Southeast public-sector partner channel · 2023)
- 139 tests and a 16-scenario eval in my open-source agent harness (basis: MIT licensed · 16 of 16 scenarios and 54 of 54 checks passing)
- 380% of quota in 2022, after 326% in 2021; never missed in any year (basis: Government Account Executive, NY/NJ/DE · blended subscribers and monthly recurring revenue)
- Quota detail: What the percentage measures: a blend of new subscribers and monthly recurring revenue, as a Government Account Executive for New York, New Jersey and Delaware.

### How I lead people and partners
#### People
- New hires: Mentored and trained the six external hires on my channel team, and sat in on every interview. Each of them closed a deal they had found themselves within three months, on a cycle that usually takes about nine.
- Promotions: Three other people I mentored were promoted. One of them is a senior manager now.
- Coaching: Took a rep who was missing quota and coached them to 115% in a quarter. They stayed above quota for the rest of 2023.
- Training: More than 1,000 dealer training attendances, including an 800-person national session for one dealer in 2019. From 2021 to 2023 I ran the training calls for the Northeast's most senior sellers: how to research a customer, how to find the data a deal turns on, and how to make T-Mobile's case to that customer instead of a generic one.
#### Decisions and money
- Partner decisions: I owned the public-sector partner channel for the Southeast from 2023 to 2025. On every deal the call on the partner was mine: whether it met the bar, whether it was the right partner for that customer, and whether we paid partner compensation.
- Forecast: I ran the pipeline and forecast reviews for our partners and for a direct sales team of about 75 in Southeast SLED.
- Budget: In the dealer channel I managed $125K a quarter in MDF and partner budgets, recommending the allocations my manager approved. Along the way I coached a dealer principal and four subagents while a Metro by T-Mobile territory grew from 8 to 20 locations.
- Business case: Today I manage resale strategy and partner relationships for systems integrators and connected solutions providers on T-Mobile's national channel team. The business case I wrote for a reseller marketplace was adopted by leadership, and the product team is building it.
#### Systems others adopted
- Prospecting system: The prospecting system I built for my territory got picked up by other regions and by the SDR team, and they still run it.
- ECF need ranking: When the Emergency Connectivity Fund arrived with no list of who needed help most, I ranked the districts by need from public free and reduced-price lunch data. The whole Northeast team worked from that list.
- Partner plan: The partner plan I wrote for the Southeast public-sector channel was adopted by two other regions.
- Targeting plans: I built the account-prioritization and enablement plans behind 310 new accounts and 600 department expansions in Southeast SLED in 2023.
- Public-sector repository: I put together a national public-sector database: every municipality with its county and independent agencies, how many people each employs, and how its fire, police and sheriff departments are staffed. Public-sector teams across the country used it.
#### How I develop people
I coach the individual and build the system around them, so what works for one person becomes something the whole team can run.
- Find the constraint: Before I ask anyone to do more, I find out what is actually holding the number back: activity, targeting, conversion, deal size or cycle length. Then we fix that one thing first.
- Coach to the person: On a new skill I go first, then step back as each person is ready. Feedback is specific: what happened, what they did, what it caused. Every coaching conversation ends with a goal and a next step.
- Coach the deal: In deal reviews we walk each opportunity together, where it really stands and every step left to close it.
- Build the system: When something works for one person, I turn it into a process the whole team can run. That is how the prospecting system ended up in other regions and on the SDR team.
I've managed people directly and I've led through influence. As a T-Mobile Retail Store Manager in Kansas City from 2013 to 2014 I had 13 direct reports, two of them people leaders. As a Government Account Executive I ran a 12-person contract research team and a marketing agency on an $8K budget. In the dealer and partner channels I led through influence: the dealer employees reported to the subagents, and the six channel hires were mentored, not managed. I finished T-Mobile's Leadership Development Program in 2018, the program that prepares people for senior-manager people-leader roles.

### Selected work
#### A territory system that hit three times quota two years running
- Result: 380% of quota in 2022, after 326% in 2021
- Decision: Prospecting system + cross-functional discovery + vehicle matching.
- Hard part: Three states with different vehicles and procurement cycles.
- What it resolved: Separately from quota, territory revenue grew 317% in 2022. The system is still running in other regions.
#### A reseller marketplace that lets partners bill on T-Mobile's paper
- Result: Adopted business case adopted by leadership, now being built
- Decision: Build a reseller marketplace where partners and third-party solutions bill on T-Mobile's paper.
- Hard part: Customers juggling a separate contract and bill for every partner solution.
- What it resolved: An adopted plan, now being built by T-Mobile's product team.
#### Community events in a dealer territory that grew from 8 to 20 locations
- Result: 8–9% added to the long-term customer base by the community events I funded
- Decision: Spend MDF on community events, with a local marketing plan for each neighborhood.
- Hard part: Reaching communities that general advertising misses.
- What it resolved: The events were the part of the budget we could measure, so that is where most of it went.
#### Ranking schools by need for the Emergency Connectivity Fund
- Result: 3,500+ Albany students connected, through the deal I closed during COVID-19
- Decision: Rank districts by need from public free and reduced-price lunch data.
- Hard part: New federal rules, and no list of who needed help most.
- What it resolved: A ranking the whole Northeast team used, and 3,500+ students online.
- Evidence: [City & State's 2022 Albany 40 Under 40 profile cites this work.](https://www.cityandstateny.com/power-lists/2022/07/2022-albany-40-under-40/374099/)
#### Opening a channel to municipal IT leaders
- Result: 2× a year the team now attends, after I earned T-Mobile's first invitation
- Decision: Earn access through direct relationships with the association's members.
- Hard part: A low-key, member-driven event where access comes by invitation.
- What it resolved: Face time, twice a year, with a buyer group that is hard to reach any other way.

### How I approach public-sector growth
- Need: Who has the problem? Examples: Agency websites, scanned PDFs and meeting recordings; Public free and reduced-price lunch data (Emergency Connectivity Fund).
- Funding: What money is behind it? Examples: E-Rate for on-campus broadband (it stopped covering hotspots and school-bus Wi-Fi in 2025); The Emergency Connectivity Fund, while it ran (2021–2024); Federal public-safety grant programs, such as COPS; State grant funding for public safety and municipalities in New York and New Jersey.
- Vehicle: How can they buy? Examples: NYS OGS contracts; New Jersey state contract; NASPO ValuePoint; Sourcewell; NPP; GSA.
- Route: Who can sell it to them? Examples: North Carolina state-contract riders through listed resellers; Partner teaming with integrators and resellers.
- Award: How does it close? Examples: RFP coordination; SLED procurement cycles; MEDDICC qualification in Salesforce.
My methodology: the first 90 days of a public-sector motion: This is how I would start a public-sector motion. Three phases, and each one ends with something you can inspect and a test of whether it worked. Partners get early milestones for customer access, contract eligibility and deal progress, and at each review I decide what to keep, change or stop.
- Days 1–30, Map how the buyers can buy. Output: A buying-path map for priority states and a shared definition of qualified pipeline. How I judge it: The team agrees which accounts come first, and why.
- Days 31–60, Build the partner route. Output: Two or three partner bets, each with a contract route and a target account list. How I judge it: Partners register their first deals on those routes.
- Days 61–90, Run the play and make it repeatable. Output: A written playbook and a weekly pipeline review the team owns. How I judge it: The team runs the review without me, and funded-need pipeline is growing.

### What people I've worked for say (Quoted word for word from public recommendations on LinkedIn. https://www.linkedin.com/in/mohelmy1/details/recommendations/)
- "I had the pleasure of managing Mo Helmy during our time together at T-Mobile, and I can say without hesitation that he is one of the most driven, innovative, and results oriented leaders I've worked with in my career." (Gary Khatcherian, Sales Manager · Channel Sales Leader; Managed Mo directly · June 30, 2025)
- "I wholeheartedly recommend Mo Helmy for any leadership role in channel sales. Mo's leadership style inspires confidence and fosters a collaborative spirit among team members." (Dave Bolton, Sr. Manager, T-Mobile for Government (DC, MD & VA); Senior to Mo · May 10, 2024)
- "Mo is a strategic planner, does his research, and attacks the market like nobody else. Because of his immediate impact while on my team, Mo earned a promotion to a role where he could significantly impact T-Mobile's Government sales." (Brett Olbrys, MPA, Executive Sales & Operations Leader; Managed Mo directly · December 1, 2022)
- "Mo is one of the most intelligent, driven, and selfless people I have ever met. It's not about him. It's about seeing his channel and the people he cares about succeed." (David Piccolo, Senior Leader, T-Mobile for Government (NY & NJ); Managed Mo directly · October 8, 2022)

### Career (most recent first)
- 2025–present: Senior National Connected Solutions Provider Executive, T-Mobile · senior manager level. Manage resale strategy and partner relationships for systems integrators and connected solutions providers on the national channel team. Wrote the business case for a reseller marketplace; leadership adopted it, and the product team is building it.
- 2023–2025: National Partner Channel Executive, T-Mobile · Southeast SLED. Owned the public-sector partner channel for the Southeast. No. 1 of 91 in T-Mobile's partner channel across every segment in 2023, Annual Winners Circle, and two other regions adopted my partner plan. Built the plans behind 310 new accounts and 600 department expansions in 2023 for a direct sales team of about 75, made the partner and compensation call on each deal, and mentored and trained six new hires. (Promoted from Government Account Executive)
- 2021–2023: Government Account Executive, T-Mobile · New York / New Jersey / Delaware. 326% of quota in 2021 and 380% in 2022, with 191 new customers. Built the prospecting system other regions and the SDR team adopted. Ran a 12-person contract research team and a marketing agency. (Promoted twice: Government Account Manager → Associate Government Account Executive → Government Account Executive)
- 2020–2021: Government Account Manager, T-Mobile. Carried government, education, public-health and nonprofit accounts.
- 2016–2020: Metro Development Manager, T-Mobile · Metro by T-Mobile dealer channel. The territory grew from 8 to 20 locations while I was there. I managed $125K a quarter in MDF and partner budgets, and my dealer trainings drew more than 1,000 attendances. (Role later titled Dealer Development Manager)
- 2014–2015: Business Manager, Human Factors International. Looked after 25+ multinational accounts buying UX services, and negotiated the master service agreements.
- 2013–2014: Retail Store Manager, T-Mobile · Kansas City, Missouri. Ran a store with a team of 13, two of them people leaders.
### Recognition
- T-Mobile Annual Winners Circle (partner channel)
- Director's Award 2021 & 2022
- City & State Albany 40 Under 40 (2022)
- Promoted 3 times in 3 years
- T-Mobile Leadership Development Program (2018)
- Press: [City & State Albany 40 Under 40 (2022)](https://www.cityandstateny.com/power-lists/2022/07/2022-albany-40-under-40/374099/)
- Education: B.S. Business, University of Phoenix (2025). Languages: English, Arabic.

## Municipal Intel (independent work)
Independent work, built on my own time from public records.
An independent research and engineering project that reads municipal documents, public meeting records and local government budgets at scale and shows what a municipality is planning, funding and leaving unfunded, with the source and the date on every line, across 1.9M+ public funding records.
Research software that reads what a municipality publishes, budgets, council minutes, meeting recordings, grant awards and announcements, and shows its strategic goals, what it is funding and what it is not. It joins what an agency needs, the money it has and the people who decide into one cited record, for anyone who wants to understand what their municipality is doing.
What a record should tell you: One record puts an agency's stated need, the money behind it and the people who decide into a single view, with the source and date on every line, so whoever is reading can check the evidence before acting on it. A funding award, a budget line and a grant the agency might qualify for are three different things; the record keeps them apart, and it says whether a connection is documented or inferred.
Design rule: Every claim points to a public source and a date. What you see is the raw evidence, dated: name, role, department, source. There are no confidence scores and no fit scores, because a score hides the reasoning you need to check.
Problem: What a municipality is planning is public, but it is scattered across thousands of agency websites, scanned PDFs, meeting recordings and dozens of funding programs, and most people only find out once a decision has been announced.
Why it's hard: The sources come in 100+ different formats, the records don't share identifiers, a lot of it is scanned paper or audio that has to be read or transcribed before it means anything, and a whole state has to be rebuildable on demand. I built it end to end, with AI coding agents writing most of the code under my direction.
- 1.9M+ funding records (basis: from 100+ sources)
- 100+ sources (basis: in different formats)
- 2,000+ municipalities (basis: seeded across seven states: the six New England states plus Pennsylvania)
- 500+ concurrent workers (basis: local and cloud capacity; a whole state seeds in a single run)
These figures count funding records, sources and seeded municipalities, not users or customers, and seeded is not the same as fully processed.
Stack: Python and FastAPI on the back end, PostgreSQL with pgvector (managed in production), Next.js in front. Search works by meaning and by keyword at the same time, and every result carries where it came from. Each workspace can see only its own rows, and operators get a read-only MCP server.
AI coding agents built most of it, under written architecture decisions and verification gates. The simplest of those gates: every database write gets read back before the job counts as done.
My contribution: The concept, the requirements and the architecture are mine, and so are the checks that decide whether a job's output is good enough. AI coding agents did most of the implementation under my direction. What I brought that an engineering team alone would not have is years of working with government: I know what a reader needs to see in a record, and what a wrong inference costs them.
Model decisions (local, hosted, or none):
- Ingesting 1.9M+ records: No LLM. Why: Must be deterministic, auditable and cheap to re-run. Effect: The full dataset re-runs the same way every time.
- Transcription and speaker diarization: Local GPU. Why: Per-minute API pricing adds up at volume. Effect: No per-minute API fees.
- Agent reasoning: Hosted + local. Why: A hosted frontier model does the thinking: it plans, reviews, edits and evaluates. Local models do the volume work, inside harnesses, evals and tests that run on my own machine. Effect: Frontier model for planning and review; local models for the volume.
- Embeddings: Full precision. Why: I don't trade retrieval quality for storage savings. Effect: Used for every embedding.
- Statewide runs: 500+ workers. Why: Serial runs are too slow for statewide coverage. Effect: A whole state seeds in one run.

## Applied AI: Applied AI systems
I build agentic AI systems for public-record research, software work and creative production. I use frontier models where judgment matters, open-weight models on my own hardware where volume matters, and plain code where the result has to repeat. Each system below gets the problem, the decisions, what came of it, and what it does not prove.
My role: I decide what gets built and how, direct the agents that build it, and decide what evidence is enough to call it done. AI coding agents write most of the code. I run them the way I ran a sales territory: written scope, acceptance criteria frozen before work starts, software rather than a model checking the result, and a receipt for every decision.
### What runs where
Most of the craft is matching the model to the job and keeping the token spend sane. This is the current line-up.
- Planning, review and judging the work: Claude, hosted; OpenAI and Gemini inside Municipal Intel. The calls that need the best reasoning, each with a receipt.
- The volume: workers, finders, summarizers: Qwen 3.8 27B on vLLM, quantized to NVFP4, on my own hardware; Kimi K3, GLM and DeepSeek hosted when I need more lanes. An open-weight model on my own GPU does the bulk work without a cloud bill, a couple at a time within the memory budget, and I reach for hosted models only when I need more lanes.
- Reading pages and pulling fields out of them: Small open-weight models through Ollama: Qwen 3.6, Gemma 4, NuExtract, ReaderLM. Structured output from messy sources without a frontier bill.
- Pictures, video, music and 3D: Flux 2 for stills; MiniMax H3, Wan 2.2 and LTX 2.3 for video; MiniMax Music 3 and ACE-Step for music; Trellis 2 for 3D, all in ComfyUI. No per-image fees, and the source art never leaves the building.
- Hearing and reading the records: Whisper and pyannote for speech; OnnxTR and Tesseract for scanned pages. Per-minute pricing adds up at volume; these run on my own GPU.
- A second opinion on a diff: Jev, a typed-judgment model. Answers in under a second and never decides; 178 cases showed why.
- Ingestion, verification and the word done: No model; deterministic code. The result has to repeat, and software has to be the one that signs off.
### The ecosystem I work in
In this field people want to know what you have actually run, so here are the names, each with where I use it.
- Hosted models: Claude (planning, review and judging the work in my agent harness); OpenAI (Municipal Intel's pipeline and embeddings); Gemini (Municipal Intel's pipeline); DeepSeek (Municipal Intel and the coding lanes); Kimi K3 and GLM (the coding lanes); Muse (Meta) (in the rotation when I compare answers across labs)
- Models I run myself: Qwen 3.8 27B (the workhorse: vLLM, quantized to NVFP4, routed through a lane router with rollback); Qwen 3.6, Gemma 4, NuExtract, ReaderLM (small models through Ollama for reading pages and pulling fields out); vLLM and llama.cpp (the engines, with two fixes submitted upstream to vLLM)
- Images, video, sound and 3D (ComfyUI, mostly for Vault 32 Interactive): Flux 2 and Flux 2 Klein (stills, with models trained on the studio's own art); MiniMax H3 and Wan 2.2 and LTX 2.3 (video); MiniMax Music 3 and ACE-Step (music); MMAudio (sound to picture); Qwen Image, HiDream, Ideogram 4, Krea 2 (other image models in rotation); Trellis 2 and Hunyuan 3D (3D from images)
- Speech and documents: Whisper and pyannote (transcription and speaker diarization on my own GPU); OnnxTR and Tesseract (OCR for scanned records); ElevenLabs and Twilio (voice and messaging)
- Retrieval: Qwen3-Embedding-8B and Qwen3-Reranker (turning text into searchable vectors and ordering the results, on my own hardware); Cohere rerank and OpenAI embeddings (the hosted alternatives, measured against the local ones); PostgreSQL with pgvector, and Pinecone (the vector stores; hybrid search with provenance on every chunk); MCP (the standard connector agents use to reach the search, read-only)
- Infrastructure: Cloudflare (Workers and R2); Neon, Supabase, Railway, Vercel (databases, deployment, and the cloud capacity behind Municipal Intel's 500+ ingest workers); Firecrawl (crawling public sources); GitHub Actions (the checks nothing lands without)
- Verification: A software verifier (signs its checks against the exact commit); Frozen, hash-locked evals (16 scenarios, 54 checks); Jev (TypeSafe) (a model that answers typed questions with a probability, used as an advisor only)
### Model operations: keeping local model services usable
Changing a model or sharing one GPU can interrupt everyone's work. These three systems make changes deliberate, failures visible and limits fair.
#### Swapping a live local model without losing service
- Problem: Swapping the model a team depends on can take the whole service down, with no clean way back.
- Decision: Every change is a dry run first, and writes a receipt before it acts.
- Decision: A process is stopped only when its id, start time and command line all match. Nothing is killed by name.
- Decision: An engine must report idle three times before it is stopped.
- Decision: Health is read from live metrics, after a hung engine kept answering "healthy".
- Outcome: After a swap, given a very long document, it still finds a fact planted 139,994, 179,978, 219,962 or 234,914 tokens in. Cold start to ready fell from 5:13 to between 3:45 and 4:04.
- Caveat: Measured on one machine, in evaluation; not yet signed off as accepted.
#### Fixing the inference engine itself
- Problem: On a hybrid model, very long prompts crashed the engine or re-ran from scratch in a loop.
- Decision: Traced it to the root cause: two parts of the engine disagreed on how large a block of memory was, so very long prompts crashed it or looped.
- Decision: A request that can never fit now ends with a clear error instead of looping.
- Outcome: A 250K-token request that extends a 140K one now finishes in 11.5 seconds instead of 25.6. I submitted both fixes upstream to vLLM.
- Caveat: Both pull requests are open and awaiting review, not merged.
- [vLLM pull request 56026](https://github.com/vllm-project/vllm/pull/56026)
- [vLLM pull request 56028](https://github.com/vllm-project/vllm/pull/56028)
#### Sharing one GPU without one user slowing everyone
- Problem: Put several people on one model server and they evict each other's cached work, and everyone gets slower.
- Decision: Each key gets its own model list, request rate, parallel limit and prompt-size cap.
- Decision: Before maintenance the gateway pauses new work and drains what's running.
- Decision: If a model is down, callers get a clear error, never a silent switch to a different model.
- Outcome: The prompt-size cap came from a measurement: one user's long prompts hit the cache on 3 of 45 requests, and output speed fell to 42.8 tokens a second against a healthy 196 to 250.
- Caveat: Serves a handful of people, not a fleet, and nobody pays for it.
### Agent workflows: keeping agent work in scope and checking completion
Long agent runs stall, widen their own scope, or report work as done before it is proven. These three systems separate implementation from verification and write down every hand-off.
#### Keeping long agent runs alive without a babysitter
- Problem: Agent sessions on local models stall, loop on one file, or just sit there waiting for someone to type "continue".
- Decision: Re-prompt only sessions with work left to do, and back off: one blocked session was re-prompted 988 times at 39 s before the brake existed.
- Decision: At most two subagents at once, because the model server's memory holds about two live requests. Extra calls wait in line; they never fail.
- Decision: Nudges go into tool output, never the system prompt, so the model's cached work stays usable and nothing is recomputed.
- Decision: Agents can't rewrite their own configuration.
- Outcome: Published as open source under the MIT license: six plugins, 139 tests and a frozen 16-scenario eval, all passing. In the eval, a blocked task got 8 re-prompts over two simulated hours instead of about 184.
- Caveat: It was built for one person's setup. In the published version the machine-specific parts are configuration, not code.
- [opencode-agent-harness on GitHub](https://github.com/MoHelmyNY/opencode-agent-harness)
#### Handing work from agent to agent without losing it
- Problem: Subagents lose context at every hand-off, widen their own scope, and stop mid-task with nothing written down.
- Decision: Every hand-off is a written artifact: a scope brief, a frozen contract, a commit with notes, a signed receipt. The next agent reads that, not the chat.
- Decision: Each role gets only the permissions its job needs. The finder, the contract judge and the reviewers can read but can't run or edit anything.
- Decision: Delegation is one level deep, from a list of 20 named subagents. A subagent can't start another.
- Decision: A worker that runs out of steps saves a checkpoint for one continuation. If it stops without notes, a summarizer writes what was tried and ruled out, so the next attempt doesn't repeat it.
- Outcome: The replay below is one ticket through this hand-off: the worker said "done" twice and was turned back 5 times before the merge. The two-at-a-time cap ships in the open-source harness with 9 tests, including first-come, first-served order for calls that wait.
- Caveat: Not yet measured: whether splitting the work across roles beats one strong agent.
- [The subagent cap on GitHub](https://github.com/MoHelmyNY/opencode-agent-harness/blob/main/plugins/subagent-cap.ts)
#### Agents that never grade their own work
- Problem: Left to itself, an agent will report work as done that it hasn't proven.
- Decision: The verifier is software, not a model. It signs its checks against the exact commit.
- Decision: The acceptance criteria freeze once written.
- Decision: A fresh-context reviewer reads the work, then an independent judge decides the merge.
- Decision: "Passed", "approved" and "pull request opened" are three different states. None of them means done.
- Outcome: The chain's final validation passed 72 tests.
- Caveat: Whether the chain beats a single agent is not proven yet, and its own plan says so.
### Retrieval: documentation an agent can trace to its source
Coding agents answer from stale memory. This system gives them the documentation for the version in use, with the source on every chunk.
#### Search that cites the version it came from
- Problem: Coding agents answer from memory, which is often stale, instead of from the documentation for the version you actually run.
- Decision: One search matches identifiers, meaning and keywords at the same time, and a reranker puts the best evidence first.
- Decision: Every chunk carries its source, version, license and commit.
- Decision: Test sets are frozen and identified by hash, so a change can't move the goalposts without being noticed.
- Decision: "Not found" never means "does not exist".
- Outcome: 1.93M passages from 239 sources, served to agents over MCP. Adding the reranker raised the share of test questions answered with the right evidence from 78% to 89%, with a median search time of 244 ms over 269 searches.
- Caveat: Self-measured on my own test set.
### Harness scorecard
A 2026 study of 97 open-source agent harnesses found 60% had neither tests nor evals; 7 had both. Here is mine, scored the same way. Most agents stop the moment nobody is watching. Mine notice the stall and keep going, with brakes on how often they try.
- Disclosed harness: Yes. Published, MIT licensed, so anyone can see what runs around the model (https://github.com/MoHelmyNY/opencode-agent-harness)
- Tests: 139 passing. Code checks that each plugin still works (https://github.com/MoHelmyNY/opencode-agent-harness/tree/main/tests)
- Evals: 16 of 16, 54 of 54 checks. A fixed set of 16 end-to-end scenarios, locked by a checksum so it cannot drift, scored (https://github.com/MoHelmyNY/opencode-agent-harness/tree/main/evals)
- Evals that can fail: Checked. Two behaviors broken on purpose; each time, only the matching scenario failed
- Software decides "done": Yes. A verifier signs its checks against the exact commit; no model grades its own work (https://www.mohamedhelmy.com/applied-ai#replay)
- Receipts: Yes. Every continuation, backoff and verdict writes a record I can re-derive numbers from
- A human can watch: Yes. Every agent session runs where I can see it live
What it doesn't prove:
- The eval simulates the model and client: it scores how the harness reacts, not gains with a live model.
- One gap it found is still open: under the current OpenCode, a repeated call is refused but never escalates to an abort.
- One author and one machine. The study only counted harnesses with several authors, so this one would not have been in its sample.
Study: [The study: State of AI Harness Engineering 2026](https://marmelab.com/blog/2026/09/24/the-state-of-ai-harness-engineering-2026.html)
### How I decide where AI belongs
- No model where the result must repeat: Bulk data ingestion runs without an LLM, so the same input gives the same output, and re-runs are cheap.
- Software verifies; models propose: A check that decides whether work is done is code, not another model's opinion.
- Dry run first, receipt always: Anything that changes a running system says what it will do, records what it did, and can be rolled back.
- Fail loudly: A clear error beats a silent fallback to a different model or a guess.
- No self-written repair code: When a scraper breaks, an agent may re-route and propose a fix. It never writes and runs new code on its own.
- Numbers come from receipts: If I can't re-derive a figure from a log or a test, I don't use it.
- How I build: I work with hosted frontier models for planning and review, and local models on my own GPU for the volume. Everything runs where I can watch it, and every change leaves a record.
### In the open
The code that can be public is public, under the MIT license, so anyone can read what runs around the models.
- [opencode-agent-harness](https://github.com/MoHelmyNY/opencode-agent-harness): six OpenCode plugins for unattended coding-agent loops on local models, with 139 tests and a frozen 16-scenario eval
- [The subagent cap](https://github.com/MoHelmyNY/opencode-agent-harness/blob/main/plugins/subagent-cap.ts): two at a time on a local model, the third waits its turn; nine tests, first come first served
- [vLLM pull request 56026](https://github.com/vllm-project/vllm/pull/56026): the memory block-size fix for very long prompts on a hybrid model
- [vLLM pull request 56028](https://github.com/vllm-project/vllm/pull/56028): a request that can never fit now ends with a clear error instead of looping
### Vault 32 Interactive (Contributor · game studio)
Vault 32 Interactive, a family-owned independent studio building games in Unreal Engine 5: I built the pipeline that makes its images, video, sound and animation on local inference, and the fail-over for its agents. (https://overtenemy.com/pages/vault-32-interactive)
- I built the pipeline the studio makes its images, video, sound and animation with, on local inference, so the art is made in-house, there are no per-image API fees, and no source art ever leaves the building.
- I'm training image and video models on the game's own art, so what they generate matches the look the studio already has.
- When a model provider goes down or rate-limits us, the agents fail over on their own, with circuit breakers and a cap on retries. Hand-offs between agents are checked, so an empty plan or a missing reviewer verdict stops things early instead of late. And whoever is operating can see stuck sessions live.

## Notes
- [Authority before action: governing agents like a sales organization](https://www.mohamedhelmy.com/notes/authority-before-action) (2026-10-02): Most agent governance reads like an IT policy. The better model is one every sales leader already runs: territories, qualification, pipeline review and approval levels.
- [I tested whether a System One model could predict my code judge. It couldn't, yet.](https://www.mohamedhelmy.com/notes/a-system-one-model-in-the-chain) (2026-10-02): Seven cases looked like a win; 178 said otherwise. What a typed-judgment model is good for in an agent chain, and what it isn't.
- [Reading notes: the harness is the product](https://www.mohamedhelmy.com/notes/reading-notes-fall-2026) (2026-10-02): Five papers from this year that say what practitioners have been finding: the system around the model matters as much as the model.
- [988 re-prompts: why my agent loop needed brakes](https://www.mohamedhelmy.com/notes/988-re-prompts) (2026-09-15): A plugin that keeps agents working is only useful if it can tell work from motion.
- [The timestamp that froze the cache](https://www.mohamedhelmy.com/notes/the-timestamp-that-froze-the-cache) (2026-09-11): With long prompts, the cheapest token is the one you don't change.
- [The engine that said it was healthy](https://www.mohamedhelmy.com/notes/the-engine-that-said-it-was-healthy) (2026-09-10): A health check tells you the door opens. It doesn't tell you anyone's home.
- Feed: https://www.mohamedhelmy.com/notes/feed.xml

## Contact
If you want to talk about public-sector growth, partner strategy or applied AI, I'd like to hear from you.
- [Email](mailto:mohamedomarhelmy@gmail.com): mohamedomarhelmy@gmail.com
- [LinkedIn](https://www.linkedin.com/in/mohelmy1): linkedin.com/in/mohelmy1
- [GitHub](https://github.com/MoHelmyNY): github.com/MoHelmyNY
- [X](https://x.com/MoHelmyNY): @MoHelmyNY
- [Website](https://www.mohamedhelmy.com/): full profile with the basis for each figure

## Docs
- [Public Sector GTM & Partner Leadership résumé](https://www.mohamedhelmy.com/resume/Mo-Helmy-Resume.pdf): PDF, two pages
- [Applied AI Solutions & Enablement résumé](https://www.mohamedhelmy.com/resume/Mo-Helmy-Resume-AI-Solutions.pdf): PDF, two pages
- [Structured profile](https://www.mohamedhelmy.com/): schema.org ProfilePage and Person JSON-LD on the home page