Autonomous Product Discovery Without an AI Manager Hierarchy
- product
- architecture
- agents
- discovery
Norditco ships several products from one lean hybrid ops setup. Coding agents already cover product ownership, adversarial critique, investment review, growth, analytics, and security. The temptation, once that roster exists, is to stack another layer of AI managers on top and let them argue until something looks like a roadmap.
That design fails in two predictable ways. First, agreement between language models is not market demand. Second, discovery work starts to compete with Product Owner for authority over the same board, so the work board fills with agent chatter while shipped products still need maintenance.
This post describes the architecture we built instead: an autonomous product-discovery workflow that reuses the agents we already have, puts a deterministic controller in charge of execution state, and treats every launch as training data for the next Scout cycle.
The goal is a learning loop
The system is not trying to prove that an idea is brilliant in chat. It is trying to:
- Discover a small number of evidence-backed micro-product opportunities
- Build the smallest experiment that can fail in public
- Measure real behaviour (starts, completions, referrals)
- Feed those results back into how Scout ranks sources and ideas
Polish belongs on the experiment, not on the concept, until the market has spoken.
Roles, reused rather than reinvented
| Role | Existing skill | Job in discovery |
|---|---|---|
| Scout | discovery Scout agent | Market signals → Opportunity dossiers with cited evidence |
| Orchestrator | discovery orchestrator + TypeScript CLI | State, scorecard, stopping rules, board and wiki write gates |
| Investor | investment review agent | Demand and distribution, one independent review |
| Critic | adversarial critique agent | Disprove the opportunity, one review, at most one revise |
| Product Owner | product-owner agent | After approval only: bound the 2–4 hour MVP |
| Builder | builder agent | Implement from templates, open a PR |
| Growth / Analytics | growth and analytics agents | Distribution and readout on launched experiments |
Product Owner remains the human-triggered owner of company roadmap work. Discovery does not replace that path. It feeds tested opportunities into it.
Opportunity contract
Every agent speaks the same JSON shape. A valid opportunity includes:
- One-sentence concept
- Audience and curiosity / share hook
- Evidence items with URL, observation date, and a reproducible query
- MVP plan capped at two to four hours on the existing stack
- Organic distribution channel and a real referral mechanism
- Success metric, kill rule, and failure modes
Without evidence and a reproducible query, the dossier never enters evaluation. That rule is what keeps Scout from inventing twenty “interesting app ideas” with no way to check them tomorrow.
Hard constraints sit above any weighted score. Hours over four, missing metrics, or an MVP that sneaks in accounts and a heavy backend fail the gate even if the narrative sounds good.
Scorecard as a hypothesis, not a prophecy
Initial weights (tunable after launches):
| Dimension | Weight |
|---|---|
| Shareability and curiosity | 25% |
| Feasibility within hours | 25% |
| Evidence of audience demand | 20% |
| Organic distribution potential | 15% |
| Operational cost and risk | 10% |
| Portfolio fit | 5% |
These weights are starting hypotheses. Only launch outcomes can improve them. The orchestrator stores results (visits, starts, completions, challenge-link opens, referred visitors versus share-button clicks) and exposes source-success stats so Scout expands categories carefully instead of scraping the entire internet every morning.
Decision protocol instead of debate
Investor and Critic run independently. They do not negotiate until one changes its mind. If they disagree, they must name missing evidence. The controller allows at most one targeted revision, then forces a decision:
- reject
- investigate (one low-cost test, often a landing or thin prototype)
- approve_mvp (at most one slot per daily cycle, rank one only)
That is deliberate. Endless multi-agent debate burns tokens and creates false confidence. The market is the only durable quality signal.
Two work streams, one controller
Discovery and operations share attention on a small company. The controller runs one stream at a time and can preempt discovery when production issues are open. Growth keeps improving live Norditco products; it does not invent net-new discovery ideas. Scout owns that branch. After an experiment launches, analytics are available to both.
Boards and wiki stay readable
Autonomy without a readable board is just noise. Agents create curated artefacts only:
| Event | Artefact |
|---|---|
| Evidence complete and MVP fits 2–4h | Opportunity page on the decision wiki (parent board item only when the write gate is on) |
| Controller decides | Decision page + one Decision item on the board |
| Experiment ships or is abandoned | Experiment Report |
No board item per agent thought. Parent themes stay outcome-sized, with a hard ceiling around ten to twelve children. The human founder should understand progress in under ten minutes. Weekly digests list new decisions, experiments, and anything blocked more than three days. SaaS automation quotas are treated as a budget, not a feature to max out; CLI and scripts do the writes.
Implementation shape
We avoided a new agent framework. The durable pieces are:
- A TypeScript package under ops with the schema, state machine, SQLite opportunity store, and CLI
- Agent skills for Scout, orchestrator, and Builder, wired to investment and critique reviewers
- Static MVP templates for challenge games and interactive tools
- Laptop systemd timers (and optional laptop workflow imports) for the daily cycle and weekly digest
- Wiki publish scripts that reuse the same diagram-as-code pipeline as our other ops documentation
Heavy parallel coding agents stay off the production VPS. Preview, tests, and human launch approval remain mandatory. Branding an unpublished local app as Norditco still requires an explicit human yes.
First milestone
One discovery cycle per day that produces three evidence-backed, buildable opportunities, a ranked recommendation, and a concrete MVP plan for the winner when gates pass. Sample fixtures already exercise the path: a shareable micro-game can land in investigate when evidence gaps remain; an eight-hour extension idea with accounts fails hard constraints immediately.
Public Lab gallery
Approved micro-experiments ship to lab.norditco.com as a static gallery, not as product landings. Each game keeps its own visual world so the experiment is memorable enough to share. The first live pack includes a regional “which costs more?” receipt game and a Danish vowel chalkboard.
The host reuses the same static-product nginx path as other Norditco privacy hosts: Cloudflare DNS, nested www sync for game subdirectories, Let’s Encrypt, and a plain /privacy page. Playable sources live under ops discovery; the public tree is a curated copy without internal architecture notes. That split keeps the Lab fast to publish while the Opportunity store remains the system of record for scores and decisions.
What we are not building
We are not building a committee of AI executives. We are not polishing concepts for weeks before anyone can click. We are not letting discovery drown the Kanban board that keeps Learn Persian, Freja, Audio Player Pro, and the hub honest.
We are building a learning system. Investor and Critic improve the decision at the margin. The market decides whether the idea works. The Opportunity contract and the bounded state machine are the shared language that makes that loop safe enough to run unattended.
