Thousand Birds is a small team building open-source infrastructure for deterministic AI agents. Chidori is the runtime; Tael is the observability layer. Both are Rust binaries you can install in an afternoon and inspect at 3am.
The goal, the context, the judgment — those are yours. The bottleneck is how much you can hold at once. Agents extend your reach: delegate to AI that remembers its work, recovers from failures, and tells you exactly what it did.
Chidori and Tael are the building blocks for that — open source, deterministic, built to be boring so the agents running on top of them can afford to be interesting.
The first ninety percent of an agent demo is easy. The remaining ten — the part where it has to run unattended, recover from a flaky tool, explain a wrong answer, survive a model swap — is where most projects stall.
The reason is structural. Most agent frameworks treat each run as a fresh act of creation: a new prompt, a new chain, a new transcript. Replays cost money, traces are second-class, and a regression at 3am means re-running the whole thing and hoping it misbehaves the same way twice.
We picked a different starting point: write agents in a deterministic dialect of Python, log every side effect with content-addressed inputs and outputs, and emit the whole thing as OpenTelemetry by default. From there, the hard problems — replay, audit, debugging, evaluation — turn into ordinary engineering.
Agent code runs inside the Chidori Starlark VM. Every host function call is logged as an OpenTelemetry span and shipped over OTLP to Tael, which stores it columnar and exposes it through a JSON-first CLI. Replay is the same binary reading the same log.
Roadmap items move when the work is done, not when the quarter ends. If something here matters to you, the fastest way to influence it is to show up in Discord and tell us why.