Autonomy
autonomy_mode answers one question: who picks the next step. That is a product decision, not an accident of the model.
The story: how it hangs together. How Runtime hydrates each mode: autonomy modes.
| Mode | Who picks the next step | Typical work |
|---|---|---|
| 0 Single inference | Nobody. One model call. No tools. | Summarise, policy Q&A |
| 1 Autonomous | The model, inside a tool budget | Investigate, look up, fan out |
| 2 Deterministic | The workflow designer | Refund, KYC, any fixed pipeline |
| 3 Guided | Designer for stages; model inside an allowlist | Triage then draft, counsel-style review |
Every route still calls the model at least once. There is no domain-HTTP-only product. A mode is not a use case — documents, human review, and handoff compose on top of these four. See use cases.
What this means in the room
- Mode 0 is a governed completion. Use it when there is nothing to invoke.
- Mode 1 is a bounded investigation. Use it when the path is not known in advance, and a budget is the control.
- Mode 2 is a process. Use it when the business already knows the stages, including a human gate.
- Mode 3 is a process with a short leash inside a stage. Use it when the outer order must be fixed and the inner choice must still be named.
Two runs of Mode 1 with the same goal may take different paths. Two runs of Mode 2 must not.
Attachments, not extra modes
These sit on the route beside autonomy_mode. Combining them is a composition, not a fifth pattern.
| Attachment | When omitted | When on |
|---|---|---|
| Retrieval | No index | Pack a corpus before generate, or call a retrieve tool |
| Memory | One-shot | Session conversation and working notes |
| Tools | Mode 0, or a Mode 2 LLM-only pipeline | Published capabilities on a manifest |
| Workflow | Mode 0 and 1 | Required on Mode 2 and 3 |
| Ingress | — | Jobs or chat. Not a mode. |
Treat memory as omit versus session. Do not explode conversation × working × loop × long-term into sixteen products.
Illegal combinations
| Combination | Why |
|---|---|
| Mode 0 + manifest | Single inference never calls a domain API |
| Mode 0 + workflow | One call has no stage list |
| Mode 0 + retrieve-as-tool | That is a tool |
| Mode 1 + no manifest | The loop has nothing to call |
| Mode 1 + workflow | The model picks the path; a stage list would fight it |
| Mode 2 or 3 without workflow | The designer cannot pick the next step |
| Mode 3 with no tools | An allowlist over an empty set |
| Long-term memory without retrieve | Facts must be recalled, not stuffed into every prompt |