AI teams need discovery systems and execution systems

AI teams get worse when they try to make one workflow do two different jobs.

Discovery and execution are not the same kind of system. Discovery is supposed to widen the search, surface contradictions, compare alternatives, and keep weak signals alive long enough to inspect them. Execution is supposed to narrow the search, choose one path, assign ownership, and deliver against a commitment.

Emad Mostaque’s cathedral-and-bazaar framing in The Last Economy is useful here because it names the split cleanly. Exploration works differently from delivery. An AI team that collapses both into one loop usually gets the worst of both: shallow research and overconfident execution.

One workflow makes the team overfit

A combined loop looks efficient at first.

The model searches, summarizes, proposes a plan, opens tasks, drafts the output, and reports progress in one pass. Throughput goes up. The team feels modern. The hidden problem is that the same system that should still be testing alternatives is already rewarded for acting as if the answer is settled.

That changes behavior fast.

Contradictions start looking like delay. Source diversity starts looking inefficient. Unranked options disappear because they were not needed for the chosen draft. The operating memory keeps the decision and drops the field of alternatives that made the decision worth trusting.

Soon the team is no longer learning. It is executing the latest plausible route with better formatting.

Discovery should preserve rejected paths

Hadto’s angle is not that every team needs more brainstorming. It is that recurring research should preserve alternatives before execution commits.

A real discovery system should leave behind a visible record of:

• which sources were checked
• which competing explanations survived the first pass
• which contradictions are still open
• which path looks best for now and why
• which paths were rejected and what evidence would reopen them

That record matters because the next operator needs more than the final answer. They need to see the choice set.

Without that, the business quietly turns back into private judgment. The agent may have done the first pass instead of the founder, but the system still cannot teach a successor how the route was chosen or when to challenge it.

Execution should narrow and commit

Execution has a different burden.

Once the team decides what it is doing, the system should stop behaving like a research notebook and start behaving like an operating contract. The execution side should name the chosen path, the artifact to ship, the owner, the review gate, the evidence needed for acceptance, and the condition that would force re-entry into discovery.

That last part is easy to skip.

Many AI teams either stay in endless exploration or pretend new contradictions do not matter because delivery already started. A cleaner system gives execution a clear escalation rule: when the work hits a contradiction strong enough to threaten the premise, execution pauses and hands the issue back to discovery instead of improvising doctrine on the fly.

This is owner-making infrastructure

Hadto’s thesis is that AI should create more capable owners, not just more output.

That only happens when the system makes judgment transferable. A future owner-operator should be able to inspect how the team explored a problem, what alternatives were kept alive, why one route became the commitment, and what evidence would justify changing course. They should inherit a business that can think in public, not one that merely ships fast.

This is the practical lesson in the cathedral-and-bazaar split. The bazaar is where alternatives stay alive long enough to find something better. The cathedral is where a chosen method becomes disciplined enough to carry real work. Healthy AI teams need both, and they need the boundary to stay visible.

The standard is simple: if the same loop is responsible for both preserving alternatives and declaring the decision final, the team will eventually optimize for speed over judgment. When discovery and execution stay separate, AI can increase owner capability instead of only increasing output volume.

---

Source evidence used in this note: Emad Mostaque’s The Last Economy (source PDF), the completed reading note at ~/.hermes/notes/reading/the-last-economy.md, the associated study ledger at ~/.hermes/notes/reading/the-last-economy-study-practice.md, and existing Hadto posts reviewed on 2026-05-08 to avoid duplicating earlier notes on discovery promotion, research handoff packets, and venue-map design.