The four lanes
| Lane | What you write | Best for | Burden |
|---|---|---|---|
| Declarative agent (manifest) | Instructions + knowledge scope + action refs (JSON; Teams Toolkit/agent tooling) | Scoped assistants on Copilot's engine — fastest credible path | Lowest: Microsoft hosts, your review is scope+actions |
| Copilot Studio (low-code) | Topics, knowledge, actions visually | Maker-speed iteration, business-owned agents | Platform governance (Studio concept) |
| Custom engine agent | Your orchestration + model choice behind Copilot surfaces | When you need YOUR models/logic/stack (existing AI investments) | Highest: hosting, safety, evals — you own the engine |
| Connectors/actions only | Data pipe or API surface, no agent | Feeding the platform (previous two concepts) | Contained |
Decision heuristics that survive contact: default to declarative until a requirement forces otherwise (custom models, complex orchestration, non-M365 channels); Studio when the OWNER is a business team; custom engine when AI is your product, not your productivity tool. And remember the licensing asymmetry (taxonomy concept): custom engines reach unseated users via Chat + credits — sometimes the lane choice IS the licensing strategy.
The lifecycle that applies to every lane
Dev/test environments (Studio) or dev tenants (code lanes) → the promotion review (scope/actions/auth — the checklists of the last three concepts) → store submission + targeting (deployment concept) → Agent ID
- CA from birth (governance concept) → versioned updates re-reviewed on manifest change. Same pipeline, four on-ramps.
What to watch (proofs)
- Lane distribution in the census: your agent inventory tagged by lane — a custom-engine-heavy estate with a declarative-shaped problem list is burning engineering on hosting nobody needed.
- Dev-tenant hygiene: sideloading/dev flags OFF in prod tenant policies (the Teams curriculum's custom-app discipline — same switches).
- Time-to-promotion per lane: how long from build to governed deployment — if declarative takes as long as custom engine, your review process (not the tech) is the bottleneck.
- Eval evidence for custom engines: test-set results shipped with each promotion — the lane that owns the model owns proving it behaves.