Agent Platform

Coding agents can already write plenty of code. The hard part is trusting it: an agent reports “done”, a locally correct shortcut gets merged, later agents copy it, and the architecture slowly degrades.

The goal, from the spec: maximize architecture-approved, verified engineering output per dollar and per minute of human attention, for one developer delegating work without watching every run.

Solo, in progress. I started with the specification, then built the MVP it scopes: declarative artifacts (JSON Schemas, workflow templates, role contracts, policies) plus a small file-based control-plane CLI.

agent init installs a .agent/ scaffold and default policies into a target repo; from there the CLI creates missions, compiles workflow templates into task graphs, and runs the claim → start → submit → gate loop, with commands for evidence bundles, architecture checks, retrospectives, replayable eval cases and a tier-gated memory.

• Wrote the v3 spec: missions, workflow templates, task and role contracts, review gates, model routing, memory and fleet security.
• Built the “agent” control-plane CLI in TypeScript: it scaffolds a target repo, validates artifacts against JSON Schemas and runs the claim → start → submit → gate loop.
• Workers never certify their own work: a task passes only with an evidence bundle, independent verification, and spec, quality and architecture review.
• Routes each task by risk to a capability tier (frontier, strong, mid, cheap), mapped to models by policy instead of hard-coded vendors.

1. Workers never certify their own work

   An agent’s own “done” is the least trustworthy signal in the loop, so completion is decided elsewhere. The evidence contract’s rule: a PASS without evidence is a FAIL.

2. Tiers, not vendors

   Model catalogs change faster than anything else in the repo, so routing only ever names a tier. The active harness profile in policies/models.yaml is the one place a tier becomes a model.

3. Schemas as the source of truth

   Every artifact shape (mission, task, result, review verdict, policies…) is a JSON Schema, validated with Ajv, so agents and the CLI agree on the contract.

4. Roles as contracts

   Each pipeline stage is a role contract (inputs, outputs, acceptance bar, failure behaviour) that can be bound to any harness: the platform’s own skills, pinned MIT skill packs, or an operator’s native install.

1. 01MissionA human writes the goal and budget
2. 02CompileWorkflow template → task DAG
3. 03RouteRisk → capability tier
4. 04WorkBounded worker in a worktree
5. 05VerifyIndependent checks + evidence
6. 06Review & mergeSpec, quality, architecture gates

From a human mission to merged, verified work. Every arrow is a file on disk, not a hidden call.

TypeScript

The CLI and its tests (Node’s built-in test runner).

Node.js

JSON Schema (Ajv)

15 schemas, one per artifact shape.

Commander

The agent command surface.

YAML

Missions, workflow templates and policies humans can read and diff.

• GitHub
• README, docs/getting-started.md, docs/evidence-contract.md, policies/models.yaml

The project’s links and the public sources this write-up draws on.