Repository intelligence for coding agents
Before an agent changes your code, it should understand the system.
Putnami is building repository intelligence: revision-specific facts, human-owned intent, and executable checks that let coding agents act with more autonomy—and show why a change can be trusted.
Or follow along before the first release.
The failure
Reading every file is not understanding a system.
- The code
- Shows what exists right now — the structure, the dependencies, the current shape.
- The people
- Carry why it exists and what must not change — intent that was never fully written down.
- The checks
- Tests and runtime evidence show whether a claim about the system still holds.
- The agents
- Reconstruct all of this imperfectly, from scratch, in every session.
The contract
Autonomy grows inside an explicit contract, not around it.
- Goal
- Intent & constraints
- Repository facts
- Proposed change
- Deterministic checks
- Evidence-backed result
The product wedge
Understand once. Reuse across agents.
Putnami aims to make repository understanding shared, revision-specific, and inspectable — instead of every agent session rebuilding an approximate model of the same system from scratch.
- Extracted fact
- A deterministic observation from code, config, schemas, or Git — parser or checker output, not opinion.
- Human claim
- A statement of intended truth from docs, ADRs, or team knowledge — useful, but unverified until checked.
- Check
- A reproducible verifier that compares a claim or contract against current facts.
- Source
- The exact file, symbol, commit, or artifact that backs a fact or a check result.
- Indexed revision
- The specific commit or working-tree state an answer is scoped to — so freshness is explicit, not assumed.
Repository intelligence is the first layer. It can support review before changes land, audit after they run, and progressively greater autonomy as the evidence earns trust.
Current status
Early, and said plainly.
- early product and practitioner research, not a finished platform
- testing the model on real repositories and real agent workflows
- not claiming complete operational understanding of any system today
- actively inviting experienced practitioners to challenge the thesis
Where to look
The working thesis is written down, with its open questions.
Repository facts, human intent, and the autonomy contract between them — laid out as a working research artifact, not a product pitch.