IAS runs Claude Code, Codex CLI, and Gemini CLI inside your own repos. Before a run starts, you can read the context pack it loads, the validators that gate it, and the approvals it needs.
Same task through IAS: the spec is already in the context pack.
The run reads it, logs what it read, and commits with the record attached.
Context loaded from docs/ias/
Spec read from context pack
Committed: a3f2b1c
Agents execute against your local checkout. The console works on run metadata and evidence; your source stays where it is.
IAS scaffolds a docs/ias/ directory into your repo. Context packs, validator definitions, and approval rules live in version control, next to the code they govern.
ias start brings the local runtime online. It claims queued jobs and executes against your local checkout, with progress streaming in the CLI.
Deterministic checks run on every step, the same rules every time. Approval gates open where your policy puts them, and only there.
Every action lands as a commit with a real SHA. Any line of agent-written code traces back to the intent that produced it.
Type an intent like “migrate the payments service to the new API contract”. IAS refines it into a spec with a readiness score, attaches the right context pack, and sets validators and approval gates before any agent runs. Step through the flow below.
* Simulated workflow for demonstration purposes
Every run loads a named pack: coding standards, compliance rules, domain knowledge. Pinned items plus declared sources, so two runs of the same job read the same thing.
Policy lives in the repo, runs happen on your machine, and oversight happens in one place.
Context packs, validator definitions, and approval rules scaffolded into your codebase, versioned with the code they govern.
ias start claims queued work, loads the context pack, runs the agent against your checkout, and commits the result.
$ ias bootstrap ✓ Scaffolded docs/ias/ directory ✓ Loaded context packs (3 files) ✓ Applied quality gates $ ias claim --queue=sprint-42 ⟳ Claimed: PROJ-142 Auth middleware ⟳ Executing against local checkout... █
Watch runs in progress, answer blocking questions, and pull the evidence trail for any change.
Agents don't stall on a missing decision. They file the question, stub the gap, and keep moving. You answer from the console when you're ready.
Work runs through role passes: PM scoping, UX validation, implementation review. Deterministic validators run alongside, the same rules on every run.
State lives in repo files, not chat history. Any agent, on any of the three models, can pick up a half-finished job, read the state, and continue.
Request a walkthrough and we'll run IAS against a real repo: an intent in, validators and approval gates on, a commit out with its evidence trail attached. Prefer to read first? The docs cover the full run lifecycle.