Agent systems · for engineering and operations teams

AI agents that run on their own and still tell the truth.

I design and build multi-agent harnesses: an orchestrator that plans, workers that gather evidence in parallel, reviewers that try to disprove every result, and gates that stop an agent until the right checks have run.

See packages Team setup from $4,000
RUNaudit · 3 apps · parallel
Plan
Orchestrator splits each app into 3 review areas: data, server, frontend.
Work
9 reviewer agents read code in parallel and return findings with file:line evidence.
Check
Each reviewer re-checks later migrations and other guards before reporting.
Verify
Top findings confirmed against source by hand. Unproven claims dropped.
Output
3 client-ready reports in one working session.

A real run from October 2026: the setup behind my launch audits.

The problem

Agents fail in the same four ways

They drift

On long tasks the agent slowly forgets what the project actually means and starts solving a different problem.

They lose the plan

When the context fills up and gets compacted, decisions and progress vanish.

They grade their own work

The same agent that wrote the code says the code is fine.

They report success that didn't happen

"All tests pass" from tests that can't fail, or a zero search result read as proof.

What I build

The harness around the model

A task goes to an orchestrator, which fans out to parallel workers; reviewers try to disprove their results; a gate checks everything ran; then a verified result. Task state and evals sit underneath.Taskyour requestOrchestratorplans and splitsWorkergathers evidenceWorkergathers evidenceWorkergathers evidence× N IN PARALLELReviewertries to disproveevery findingGatechecks ran?ResultverifiedTask state outside the contextsurvives compaction, so a run can last hours without losing the planEvalsmeasured before and after
The agent that does the work never grades it. Reviewers and gates decide what counts as done.

Hierarchical agents

Orchestrator, cheap parallel workers and independent reviewers, each with a narrow job and an evidence standard.

Long-running loops

Task state saved outside the context, checkpoints and recovery, so a run can last hours without losing the thread.

Hard gates

Hooks that block edits, commits or releases until impact checks and reviews have run. Rules, not suggestions.

Real context

MCP servers, code knowledge graphs and project invariants that outrank every other document.

Evals

Measured before and after, including honest negative results, so you know the harness helps.

Cost and safety limits

Per-run budgets, model tiering, prompt-injection boundaries and audit logs of what each agent did.

Packages

Start small, then build what pays off

Team agent setup

from $4,000
1 week
  • A governed Claude Code, Codex or Cursor setup in your repos
  • Project rules, review gates and sub-agents tuned to your codebase
  • Half-day workshop for your team
  • Two weeks of follow-up tuning

Custom agent system

from $12,000
after a $1,500 discovery
  • A multi-agent workflow built for your process: research, code review, operations, support, data extraction
  • Orchestration, evals, cost limits and monitoring included
  • Runs in your infrastructure, documented for your team
  • Discovery fee credited against the build

Ongoing

$3,000/month
monthly
  • New agents and workflows as you need them
  • Tuning when models or your codebase change
  • Review of agent runs and costs
How it works

From a call to a working system

  1. Fit call, 20 minutes. What you want agents to do, where they fail today, what "done" looks like.
  2. Discovery. I map the workflow, the data and the risks, and send a written design with a fixed price.
  3. Build. Weekly demos on your real tasks, with evals that show progress in numbers.
  4. Handover. Documentation, a team session, and limits and monitoring switched on.
Built before

Agent systems I've shipped

I build these for my own work first, then for clients. Most of the code is open source.

  • Case study: 14 AI agents build, ship and support a family of game mods: agents develop the mods, publish releases, ship weekly collection updates and support a 600-member community. 26 collection updates and 2,824 messages to players in 34 days.
  • Case study: nine AI agents, one codebase: a team of autonomous agents that builds and runs a live trading-research system. 306 commits and ~3,400 tests in eight days.
  • Featured by Akon Labs (YC S26), the company behind GitNexus, as an example of agent infrastructure built on its code graph. Its co-founder called it "really great work".
  • bearing: an open-source governance layer for Claude Code, Codex and Zed agents with hook-enforced gates, task state that survives compaction, adversarial review waves, evidence-only worker agents and an eval harness. Used on six production codebases.
  • The multi-agent audit pipeline behind my launch audits.
  • An agentic identity-resolution pipeline for a sales team: search, verify and cite evidence for every match, at $0.95 per 15-call run.
  • MCP-based agent actions inside an AI assistant for US medical billing teams.
  • Maintainer of GitNexus, an open-source tool with 48,000 GitHub stars that gives AI agents real context about code.
Questions
Which models and tools do you use?

Whatever fits the job and your stack: Claude, OpenAI and open models, Claude Code, Codex, Cursor, Zed, MCP. I don't lock you into one vendor.

Will the agents touch our production systems?

Only where you decide, and behind gates. By default agents work read-only or in staging, with every action logged.

How do we know it works?

Every build has evals on your own tasks, measured before and after. If a part of the harness doesn't help, the numbers say so and we drop it.

Do you sign an NDA?

Yes, on request. Your code and data stay in your systems wherever possible.

Book a 20-minute fit call

Tell me what you want your agents to do.

Describe the workflow and where it breaks today. I reply within one business day.

hello@vadimlysenko.com