agentic-ai-engineering/products/AGENT_PRODUCTS.md

166 lines
4.7 KiB
Markdown

# TAC Agent Products
Build products as agent-run systems, not static templates. Each product packages TAC skills, course ideas, tests, deployment, and an operating loop.
## Product Spine
Every product gets the same boring skeleton:
```text
idea -> plan -> build -> test -> verify -> signed deploy -> observe -> improve
```
Runtime pieces:
- ADW pipeline for plan/build/test/review/ship.
- Verifier Pro before ship.
- Security Foundation around tools and deploy.
- Observability events for every action.
- Autoresearch/SkillOpt loop for measured improvement.
- Agent-native memory: events, claims, evidence, state cards, action outcomes.
## Product Line
| Product | User | Agent skills bundled | Runs on agents by |
|---|---|---|---|
| Product Factory | founders/builders | ADW, orchestration, verifier, deploy gates | turning ideas into shipped agent apps |
| Security Gate | teams with dangerous agents | L3-L5 security, prompt injection defense, sandbox | blocking unsafe commands and deploys |
| Verifier Pro | anyone shipping with agents | verifier, confidence ladder, claim decomposition | checking builder work every turn |
| SkillOpt Lab | agent-system owners | autoresearch, integrity guards, mutation ledger | improving skills only when evals pass |
| CEO Board | strategy/product decisions | CEO board, validator, tracker | adversarial decision memos and logs |
| Observability Cockpit | operators | tracer, cost tracker, replay, loop detector | showing what agents actually did |
| Multi-Agent Orchestrator | engineering teams | chains, teams, P2P, domain locks | routing work to specialists |
| Task Discipline | vibe-coders | tilldone, purpose gate, progress nudges | forcing defined tasks and completion |
| Agent Memory OS | research/factory users | evidence, claims, state cards, outcomes | recalling what worked and why |
| Local Model Gateway | privacy/resilience users | model routing, safety gate, local Kimi/Ollama | using local models without deploy authority |
## MVP: Product Factory
The first sellable product should be **Product Factory** because it contains the whole TAC thesis.
### What it does
```text
User submits product idea
-> planner creates scope and acceptance checks
-> builder creates minimal app/service/agent
-> tester runs checks
-> verifier grades claims
-> deployer uses signed action API
-> observer records evidence
-> SkillOpt loop improves weak skills later
```
### Included skills
- `agent-chain` for ADW phase flow.
- `agent-team` for specialist fanout.
- `verifier-builder` for read-only review.
- `confidence-ladder` for pass/fail grading.
- `damage-control` and `l5-no-bash` for safety.
- `tool-call-tracer` and `session-replay` for observability.
- `experiment-loop`, `integrity-guard`, `median-over-best` for self-improvement.
- `tilldone` and `purpose-gate` for task discipline.
### Control plane
Use the existing engine factory:
- `git-proxy` / AI proxy: `8099` — primary signed deploy path: `/deploy`
- `agent-site`: `8084`
- `deploy-webhook`: `8098` — legacy/internal fallback; do not call directly from TAC product flows
- Forgejo: `3030`
- Qdrant: `6333`
- Prometheus/Grafana: `9090` / `3001`
Deploy must use signed actions, not raw shell.
Allowed actions:
```text
run_skill_tests
patch_skill_from_pr
restart_container
rollback_container
disable_skill
enable_skill
```
## Build Plan
### Phase 1 — Catalog
Generate `products/catalog.json` from existing product READMEs and skill catalog.
Fields:
```json
{
"id": "product-factory",
"name": "Product Factory",
"skills": ["agent-chain", "verifier-builder"],
"agents": ["planner", "builder", "reviewer", "executor"],
"checks": ["pytest", "verifier"],
"deploy_actions": ["patch_skill_from_pr"]
}
```
### Phase 2 — Runner
Create one CLI:
```bash
python products/product_factory.py "build an agent-powered product"
```
It should only:
1. create a product folder,
2. write a spec,
3. run ADW pipeline,
4. record events.
No new framework.
### Phase 3 — Evidence ledger
Add JSONL:
```text
products/.events/YYYY-MM-DD.jsonl
```
Record:
- idea received
- plan written
- checks run
- verifier result
- deploy action requested
- outcome
### Phase 4 — Dashboard
Static markdown first:
```text
products/DASHBOARD.md
```
Show product status, last check, deploy readiness, and open risks.
## Done Criteria
- `products/catalog.json` exists.
- `products/product_factory.py` creates a product spec from an idea.
- `products/.events/*.jsonl` records actions.
- `uv run --with pytest pytest -q` passes.
- No raw shell deploy path is introduced.
## What Not To Build Yet
- No custom marketplace.
- No new graph DB.
- No giant web app.
- No autonomous agent spawning.
- No local Kimi deploy authority.
Ship the file-based factory first. Add UI only after the loop works.