LAB-05
AgentFlow
AgentFlow monitors AI agent systems, detects failures and sends alerts. A TypeScript monorepo of six packages, MIT licensed, with a documentation site.
- Status
- live
- Stack
- TypeScript · Node.js · OpenTelemetry · React · Docusaurus
Known issues
- The public repository does not contain the SOMA detection engine, which is proprietary and lives in a separate private repository. What is public is the integration surface around it, meaning the adapters, the dashboard components and the documentation.
- Only the core package has zero runtime dependencies. The dashboard package depends on express, react, react-dom, ws, chokidar and express-rate-limit, so a blanket zero-dependency claim about AgentFlow would be wrong.
Clemence W. Chee built AgentFlow to watch AI agent systems while they run, catch the failures they produce, and raise an alert when one appears.
What does AgentFlow do?
AgentFlow monitors a running agent system, detects failures automatically and sends alerts.
An agent system fails differently from a conventional service. Nothing throws. The process exits zero, the trace looks ordinary, and the damage is in the content of what the agent decided. Monitoring that requires instrumentation at the agent boundary rather than at the process boundary, which is the problem AgentFlow addresses.
What is in the repository?
A TypeScript monorepo with six workspace packages: core, cli, dashboard, bi-platform, otel and python.
The core package carries no runtime dependencies at all. The dashboard package is the other extreme, pulling in express, react, react-dom, ws, chokidar and express-rate-limit, which is why the zero-dependency claim on this page is scoped to core and stops there.
Around the code sits the apparatus that decides whether anyone else can use it. Tests run under vitest, formatting and linting under biome, types under tsc -b. GitHub Actions builds the project and publishes the documentation. Dependabot tracks the dependency graph, and SECURITY.md and CONTRIBUTING.md tell a contributor what to do before opening anything.
Why is the detection engine missing from the public repository?
The SOMA engine is proprietary and sits in a separate private repository. AgentFlow publishes the integration surface around it.
That split is a decision rather than an accident of packaging. The adapters, the dashboard components and the documentation are the parts another engineer has to read in order to wire the thing into their own system, so those are public. The detection logic is the part Clemence W. Chee sells, so it is not.
Who wrote the code?
AgentFlow is built with AI assistance. The commit history carries two names: Clemence W. Chee, and Shelly, which is his second AI assistant.
He owns the architecture and the review of what lands, and he decides which parts of the system are public and which stay private. Anyone reading the history will see the second name in it, so it is stated here rather than left to be discovered.
What is the current status?
AgentFlow is live. The repository was created on 2026-03-13, it is MIT licensed, and the Docusaurus documentation site is published from it.
The documentation URL above serves the real docs and answers with HTTP 200, which is more than two of the other entries in this lab can say. It is the artefact on this site that a technical reader should open first.