Mark Feldman · Herriman, UT · Remote
Agentic Systems Engineer
I build systems where AI agents do real engineering work under real constraints — multi-agent orchestration, permission and authorization boundaries, and the review discipline that makes agent output safe to merge.
I'm looking for a role helping a team adopt agentic AI. I spent 12+ years leading engineering teams before this, and the hard part was never the tool — it was getting an organization to use a capable new thing well, and making the practice stick.
How I actually work with agents
Most “AI experience” means prompting a chat window. This is the operating practice I run instead.
Parallel dispatch, isolated worktrees
I run multiple agents concurrently, each in its own git worktree, so simultaneous work can't collide. Branches come back one at a time to be reviewed and merged deliberately.
Every agent branch gets read before it lands
Not spot-checked — read. On my current project that caught an unvalidated shell interpolation in agent-written dispatch code: a real injection vector, fixed before merge. Agents produce plausible code; plausible is not the same as correct.
Model tier matched to stakes
Cheaper, faster models for mechanical subtasks; stronger ones where judgment actually decides the outcome. Cost control that doesn't quietly degrade the work that matters.
Hard human-only boundaries
Credentials, live domains, and real spend stay with me. When an agent hits an authentication wall, the right behavior is to hand it back — not to route around it. That boundary is a design decision, not an accident.
"Done" requires evidence
A dependency once blocked real end-to-end verification on my governance project. That shipped as a documented known limitation rather than a passing gate. Adoption fails fastest when a system is trusted past what it has actually proven.
Current projects
Foundry is the operating layer I am building for repeatable, governed AI work. Cerberus is the large contract-driven build that keeps pressure-testing how agents behave when invention is forbidden.
Foundry — the operating layer for dependable AI work
General agents can reason, code, and use tools. Foundry makes a repeated job survive the chat window: the work is versioned, context is traceable, progress is checkpointed, authority is explicit, and consequential outcomes still stop at a human gate.
Five chapters are implemented through durable workflows. Chapter 5 is active: turning a proven repository-inspection procedure into a reusable Skill that carries instructions without carrying data, credentials, or authority with it.
Explore the Foundry concept and build →Judgment moves. Authority does not.
Deterministic software keeps the books, AI handles bounded judgment, and a model cannot approve its own consequential result. Permission is a validated contract, not reassuring prose in a prompt.
It is designed to outlive a provider
Hosted and local models sit behind one inference boundary. Workflows, approvals, evidence, budgets, and run history stay with Foundry instead of disappearing when the model changes.
151 tests, 149 subtests, and a stricter red gate
The current working tree passes the full suite. A recent Steward run also proved that safe execution was not enough: its answer quality failed. The evaluation now separates those claims, and a fresh end-to-end quality proof remains open.
Perimeter-Cerberus — agents under a frozen contract set
A game built against 356 design contracts that are the sole authority: no gameplay value, ID, reward or UI state may be chosen or inferred outside them. It is the closest thing I have to a controlled experiment in what agents do when guessing is forbidden — because every time one wanted to fill a gap, the gap got written down instead.
Twenty-six sprints, a shipped Windows build and Android APK, and two registers of everything the contracts got wrong. The implementation stays private — it is an unreleased game — but the engineering record is published here.
Read the project roadmap →It refused to invent a value nobody wrote
Career Rounds 1 through 3 set a minimum spawn separation of 20°, 18°, 16°. Rounds 4 and 5 state none. The obvious continuation is 14 and 12, and that is exactly the problem — it would put a spawn rule into the game under a citation that never mentions one. The generator emits zero, prints a warning, and the gap is filed as a defect awaiting an amendment.
A success that was silently wrong
A generator's pattern matched letters followed by at least one digit, so three letters-only skill keystones never matched and never became nodes. Nothing failed. The output was well-formed and 33 nodes is a perfectly plausible number. It surfaced only because a second, independent reading of the same contracts was cross-checked against the first — which is the only reason anyone found it at all.
Two of eighteen release gates met
The other sixteen are listed as pending rather than deferred out of sight. At the frozen v1 tag, all 1,569 tests pass; CI runs the same suite on Linux, Windows, and macOS. That is the baseline the second build has to preserve, not a claim that v2 already exists.
Atlas — a governance layer for multi-agent work
Atlas explored authorization and decision structure for teams of agents: a human director authorizes, an executive agent plans, specialized advisors inform hard calls, and bounded operators act. Its most important lesson now feeds Foundry: enforcement belongs in task shape and validated state, not in prompt language alone.
The implementation remains private. The public record includes the architecture decisions, phased plan, risk register, and the gaps that were deliberately left visible.
Read the Atlas design record →This site is a work sample too
Built with Claude Code, not a template — real architecture tradeoffs, real deployment friction, and real decisions about what to delegate versus keep human-controlled. The panels below are live from this repo's own GitHub API and CI, not screenshots.
The background this rests on
15+ years in C#/.NET, 12+ leading the teams that ship it. My track record is walking into organizations with no engineering practice and installing one that lasts — which is the same problem as adopting agentic AI well.
- Senior Development ManagerDental Intelligence2024 – 2026
- Software Engineer IIMelaleuca2023 – 2024
- Software Development ManagerMartin Garage Door2022 – 2023
- Software Development ManagerChartLogic / Medsphere2015 – 2022
- Senior Software EngineerEchopass / Genesys2014 – 2015
Enterprise SaaS, healthcare technology, and analytics — including a multi-tenant EHR platform under HIPAA, SOC 2, and ONC obligations.