AI Agent Orchestration Platform · View 07 of 32 · 2 · Structure
Decisions
- The worker is stateless between steps; everything that must survive is written at Commit and nothing is held in memory across a lease
- Guardrails run in-process on both sides of the model call — a network hop per check would double orchestration overhead
- The agent framework sits behind an adapter, so a LangGraph or Semantic Kernel agent runs on the same worker image
Limits enforced here
- Wall-clock timeout, token budget, cost budget and recursion depth, all from the pinned agent spec
- Tool-call validation happens before invocation, not after: schema, permission grant and argument inspection
- Structured output is enforced by schema; a non-conforming response is retried once, then escalated
Assumptions
- Content Safety prompt shields and groundedness detection are available in the execution region
- A step's work is idempotent given its step key; handlers that cannot be made idempotent must declare a compensation