A team can answer “can we build this?” and still be unable to answer “have we agreed how this should work?”
That gap is easy to miss. A working prototype makes a question feel settled because something now runs. But the prototype proves only that one path can be implemented. It does not prove that the work is correctly framed, that its constraints are shared, or that someone owns the exceptions.
An approval flow is more than its happy path
Consider an illustrative approval-routing system. The prototype accepts a request, sends it to an approver, records the response, and releases the next step after approval. The main path works. The team can see the request move from submission to decision.
What remains undecided sits around that path. Which requests need approval at all? Which source controls when policy and commercial guidance conflict? Who may override a rejection? What happens when the approver does not respond? Which evidence must be retained, and who is accountable when an exception becomes precedent?
None of those questions is solved by generating more workflow code. They require a shared account of the work: its governing constraints, decision rights, evidence requirements, authority limits, and escalation paths.
Implementation turns a defined path into working software. Registration does the prior work: it captures the rules, exceptions, judgment, and context experts carry; resolves contradictions; and assigns the evidence, decision rights, authority limits, and escalation paths that govern execution. The resulting source of truth preserves why a decision is valid and who owns what happens next, so humans and agents can build, operate, verify, and revise against the same account of the work.
Zach’s experience building Induct reinforced the point: agent speed was real, but the difficult work was making domain rules, exceptions, decision rights, and evidence requirements precise enough that the system could act without guessing. Once that layer was solid, humans and agents could move much faster; without it, AI could accelerate building the wrong thing.
As models make more execution possible, ambiguity, conflicting rules, unclear authority, and weak outcome verification become more consequential. Faster execution raises the cost of an ill-defined target.
The practical diagnostic is short:
- If the main path is unclear, define the work before building more.
- If constraints conflict, name the controlling source and decision owner.
- If exceptions are unowned, assign authority and an escalation path.
- If material claims lack support, stop release and attach evidence or narrow the claim.
- If all four are explicit, implement the next bounded step and verify it.
Faster production is useful. Before building the next bounded step, choose: resolve the work well enough to authorize it, or accept that speed will compound the ambiguity.