Kukui Bot · Engineering
A size-gated multi-agent pipeline: multi-model planning diversity, a single project manager who synthesizes the plan, a right-sized coder, and evidence-based QA. The default for real product work — not a rigid ritual for every typo.
Multi-model plan diversity → one owner synthesizes → fast default coder → independent evidence QA. Matches how a multi-provider agent platform should route work.
Full five steps for every typo burns latency and noise. Small work short-circuits. Huge work upgrades the coder — it does not force “always Grok + always four planners.”
What a normal feature or fix looks like end to end.
Clarify
Scope, options, brief. Brainstorm before a big dispatch.
Fan-out
One prompt → several strong models in parallel (often Grok, Opus, Codex, Composer).
Synthesize
Merge into one executable plan. Tradeoffs go to the human for architecture calls.
Implement
Default Grok. Auto-route to the best free developer. Keep phases short.
Final QA
Re-run tests or inspect files. An evidence gate — not a rubber stamp.
| Step | Job | Why it exists |
|---|---|---|
| PM clarifies | Owns scope, options, and the brief before anyone codes | Prevents thrash |
| Planning fan-out | Several models answer the same hard question in parallel | Beats single-model blind spots |
| PM synthesizes | One owner turns diversity into a single plan | Avoids committee coding |
| Developer implements | One coder ships against the plan | Speed with ownership |
| PM final QA | Independent check with fresh evidence | “Done” is proven, not claimed |
Best process = the pipeline plus these gates. Short-circuit small work; upgrade the coder for hard work.
Skip planning entirely. Prefer the cheapest capable model.
This is the standard product path.
Aim for short phases. Optional separate test phase. Humans decide architecture tradeoffs.
Depth is worth the wait. Not the default coder.
When no model is forced, the platform prefers this order among free developer seats — never invents a missing model.
Multiple models plan in parallel; one PM synthesizes. Avoids committee coding and single-model blind spots.
Full pipeline only when ambiguity or blast radius justifies it. Small work short-circuits.
PM re-runs tests or inspects files. Developer “done” is a claim, not a fact.
Keep dispatches short. Split implement vs test when both are non-trivial.
Coordinates, synthesizes, gates. Architecture options come with tradeoffs — the human decides.
Delegates do not bounce the server. Prefer hot-reload; full restart only when core files change.
| ✓ | Deliverables match the brief |
| ✓ | Tests or verification evidence present (fresh, not remembered) |
| ✓ | No silent scope expansion |
| ✓ | Plain-language status for the operator — not a wall of code |
| ✓ | Related code considered; neighbors not broken |
Kukui Bot is self-hosted multi-model agents with real tools, delegation, and evidence-based workflows.