Agent Workbench

Agent Workbench

Adversarial verification for AI-agent workspaces. The core path:

External Coding Agent -> Controlled Verification -> Evidence
-> Deterministic Verdict -> Immutable Receipt

**Primary deliverable: the `aw` CLI** (`aw-verify` npm package). Run a trusted recipe against a git workspace and get an immutable, evidence-bound receipt that says what was checked, what the honest verdict is, and which code state it was bound to.

Agent Workbench is an independent project and is not affiliated with or endorsed by Anthropic or OpenAI. Vendor names identify supported external tools only.

Install & use

npm install -g aw-verify          # or: node cli/build.mjs && npm link

aw init ./my-workspace            # scaffold contract.json + test/verify.test.mjs
aw verify ./my-workspace --contract ./my-workspace/contract.json   # exit 0 = VERIFIED
aw inspect ./my-workspace         # read-only report, never executes a recipe

• verify exit codes: 0 VERIFIED, 1 FAILED / INSUFFICIENT_EVIDENCE, 2 usage or environment error — drops straight into a CI gate.
• verify writes an immutable verification-receipt.json + verification-handoff.md.
• Receipts are display-safe and evidence-bound. VERIFIED is not human acceptance; AcceptanceDecision remains NOT_RECORDED.

Safety model

• The recipe is a fixed node --test run by a trusted external node.exe with an allowlisted environment; arbitrary commands are never executed.
• The Subject Snapshot is captured before/after and must be byte-identical — a test that mutates the workspace fails closed (INSUFFICIENT_EVIDENCE).
• No network, no server, no Electron, no telemetry.

Repository layout

• src/engine/ — the verification engine (pure Node; also consumed by the desktop app).
• src/shared/ — pure shared types and logic (contracts, receipts, criteria).
• src/utils/ — pure utilities.
• cli/ — the aw-verify npm package (esbuild bundle of `script