Agent frameworks solve orchestration. Nobody's solving state, recovery, or cost control.
Orchestration is 10% of the problem
Every agent framework sells you on the DAG - define nodes, wire edges, let the LLM decide. Clean abstraction, great demo. Maybe 10% of production agent work.
- The other 90%: state management, failure recovery, cost tracking, observability
- None of it is in the README
State management is the real work
30-second agent doesn't need state management. 30-minute agent does.
- Moment your agent touches external systems, you have a distributed state problem
- The question isn't "what step is it on" - it's "if this dies now, can I resume without corruption or repeated side effects?"
- Most frameworks punt entirely
Recovery requires transactional thinking
- Retry logic is table stakes
- Recovery means: what already happened, what were the downstream effects, has the world changed since the last checkpoint?
- This is transactional thinking, not orchestration thinking
- Teams that get this right come from infrastructure backgrounds, not ML
Agents don't know when to stop
- Agent with tools and a vague objective will burn tokens forever
- Budgets help (token limits, step limits, time limits) but aren't the real solution
- Real solution: tighter scoping - small focused agents that finish fast and hand off cleanly
- Composition over complexity
The boring parts are mandatory
Logging, observability, cost tracking, audit trails. Not exciting. All mandatory.
- If you can't answer "what did the agent do, why, and how much did it cost?" for every run, you're not production-ready