Choruz: An Agent Workspace That Learns From Your Work
libraryYour notes
Every published coding-agent number is measured on somebody else's tasks. Choruz, from Ant's inclusionAI org, starts from the observation that the benchmark you actually care about is the work you already did, and that the thing worth changing in response is not the model. It is an agent workspace for Codex, Claude Code and Muse that asynchronously turns your interaction traces into personalized benchmarks, then uses them to adapt the harness — the README's own line is "this adapts the harness, not the model's weights", and the selection criterion is explicitly task suitability, "not a provider model's parameter count or billing cost". That is a model–harness-pairing thesis stated as a product.
The loop runs behind your foreground work: your task plus the agent's actions plus your real feedback → background trace analysis → personalized benchmark cases → evaluation of candidate prompt and team changes → reviewed improvements applied to future tasks. Three choices give it teeth. Cases carry an input, an expected outcome, and either an exact check or an independent AI judge, with curated training, validation and held-out splits so evaluation stays separate from optimisation. The analysis and evaluation standards are held fixed while the agent's instructions and execution team are allowed to move — you cannot improve the scorecard, only the harness. And escalation is gated: prompt changes come first, and only a documented recurrence after guidance was already used unlocks changes to collaborator count, role prompts and execution order; revisions can be restored or cleared. A nice detail in the trace-labelling: "an unsuccessful exploration is not automatically an agent mistake."
The pairing claim becomes concrete in selective provider-model execution. Once you configure a smaller provider model, a validated finite-output program can complete applicable turns without native inference, with confidence and applicability checks, recorded model attribution, and abstentions or provider failures falling back to the original agent. The README refuses the obvious overclaim twice: this is "not blanket replacement of arbitrary tasks or a guaranteed cost saving", and of the demo, "no measured improvement or cost saving is claimed." Claude Code, Codex, Muse Code and webhook-driven external agents are supported by default, with plugins for Grok, Pi, OpenCode and MathCode — but native trace learning currently covers only Codex and Claude Code; Muse runs in the workspace without a learning adapter yet.
Engineering-wise it is a real workspace rather than a library wrapper: Rust plus Node.js 24 and PostgreSQL 16, with the choruz-learning, choruz-evaluation and choruz-decision crates usable independently of the UI, a published OpenAPI contract, WebSocket sync, webhook agents and Slack/Telegram bridges. It is labelled a Developer Preview whose interfaces, configuration and data formats may change incompatibly, and the background analysis runs in separate asynchronous sessions that consume the selected account's own usage. Where AntOmniEvo, Ant's other auto-evolution release of the same month, optimises a directory of tunable artifacts against an eval set you supply, Choruz generates the eval set from your traces first. Apache-2.0, ~1,024 GitHub stars and 9 forks at filing. Note that Ant publishes from both ant-research and inclusionAI on GitHub.