Frameworks move fast. Agents are going to production before the infrastructure to operate them exists. The industry spent decades building reliability primitives for services — observability, cost governance, regression testing. For the agentic stack, that work is just beginning.
We build that layer. Open source, shipped in the open as we learn.
A production agent fails mid-run. The prompt, state, and tool chain that caused it are gone.
A multi-agent workflow burns 10× its token budget because each agent caps spend locally.
You ship a fix for a regression — but without a committed test, nothing stops it from coming back silently.
Monitoring built for stateless services doesn't transfer. Testing frameworks built for deterministic code break on probabilistic outputs. Something purpose-built for agents is needed.
Record-and-replay for agent decision graphs. pip install agent-chronicle
Chronicle instruments every LLM call, tool call, and routing decision as an immutable Envelope. When something breaks in production, that incident becomes a committed regression test — replayable without a single live model call.
Run-aware token governance for multi-agent systems. pip install agent-tokenops
TokenOps governs the run, not the request. A shared ledger and control plane enforce spend policies across every agent, model call, and tool hop in a workflow — before the next LLM call executes, not after.
Chronicle and TokenOps compose: Chronicle records decision boundaries; TokenOps attaches as the cost and governance observer on live crossings.
Chronicle and TokenOps solve two problems we hit first. The agentic stack needs more:
- Memory that doesn't hallucinate past state and survives context windows
- Evaluation built for behavior, not output strings
- Retrieval that gives agents accurate, scalable perception of large codebases
- Observability at the workflow level, not the request level
We're building in that direction and publishing the thinking as we go.
| 🎥 Your Agent Failed in Prod | Why replayability beats bitwise determinism — Chronicle walkthrough |
| ✍️ Your Agent Failed in Prod | The nondeterminism problem and what record-and-replay actually gives you |
| ✍️ From Validating Code to Evaluating Behavior | Why traditional testing assumptions break for agentic systems |
| ✍️ Why Agents Forget | How memory works in AI agents, and what it takes to make them learn |
| ✍️ Beyond Prompt Caching | AST-aware indexing, multi-modal search, and codebase perception for coding agents |
| ✍️ Spec-Driven Development | Token economics and when structure helps vs. becomes tax for AI agents |
All writing and talks → theagentplane.github.io/media
Issues and PRs welcome on any of our repos.