I build autonomous AI systems for institutions where failure has legal consequences. Banks, insurers, healthcare. Then I publish what held up and what did not.
Everything on this profile is a build I shipped, not a tutorial. If a repo says it runs, it runs.
Autonomy is table stakes. The moat is everything around the agent: verification, orchestration, human oversight, grounding, and an audit trail that survives a regulator.
A series of real builds on the agentic stack. Each episode takes one layer everyone is talking about, turns it into a working system, and ships the film and the code together.
agents-orchestratedis the series index. Films land on LinkedIn.
| EP 01 |
Loops Verification is the moat. Every agentic coding tool converged on the same shape: goal in, iterate, verify each pass. Why the loop you can audit is the only loop that ships. 
|
| EP 02 |
AG-UI Human in the loop, orchestration and grounding are the moat. Agents can stream the whole interface, not just text. A private bank onboarding clients where every agent decision arrives in the browser as live, clickable UI. When the money gets serious the process pauses for a human, visibly. Runs without API keys. 
|
| EP 03 |
AP2 Cryptographic authority and governance are the moat. I gave an agent a signed $7,000 ceiling, let it buy World Cup tickets while I slept, then tried to break it. Tampering with the signed limit gets refused by cryptography. Getting close to the limit gets escalated to me by a decision table. Every hop is one auditable process. 
|
| EP 04 | Next up: model routing, and what it costs you to pick wrong. |
Reference documents on the open protocols, written from builds rather than from specs.
| Guide | On |
|---|---|
| Where Agents Actually Work | Field notes from 50+ enterprise builds |
| The Control Plane | Governing agentic AI at scale |
| Evaluating Agentic Systems | The four axis method |
| Tool Contracts | Why agents skip your tools |
| AP2 | The Agent Payments Protocol |
| AG-UI | Streaming the interface, not just the answer |
Agents fail in production because tool specs are underspecified. Paste an MCP tool spec and it scores the contract, then auto-fixes it. Deterministic first, model polish optional and local only.
Because that is the job. Anyone can get an agent to do something impressive once. The interesting question is what happens on the run where it is wrong, or slow, or down, and whether you can prove afterwards what it did and who approved it.
That is what I build, and it is what I write about.