Managed harness: what do we hand over?
Anthropic Managed Agents and the OpenAI Agents API: task state, stops and the team's control
Slide contents
1. Managed harness: what do we hand over?
Anthropic Managed Agents and the OpenAI Agents API: task state, stops and the team's control
2. Platforms run the loop; outcomes stay yours
Anthropic Managed Agents and the OpenAI Agents API in three questions
3. The review follows three questions
First why the harness was split, last the caveats to the comparison
4. A shared container becomes a pet
The first design kept the log, the harness and the sandbox together
5. Three interfaces are meant to outlive implementations
The session, the harness and the sandbox fail and get replaced separately
6. A brain without a container answers sooner
The sandbox is provisioned by a tool call once it is needed
7. OpenAI hosts the Codex harness for you
You choose the environment: none, OpenAI-hosted or your own
8. The session log outgrows the context window
Compaction is irreversible, but events can be read again
9. OpenAI saves items, not the stream
Events show live progress; items remain in the history
10. Files expire with the sandbox
History and the working directory have different lifetimes
11. Some state stays in your application
The platform keeps the session, not its link to your product
12. Both platforms keep history on their side
Your own sandbox keeps the files with you, not the history
13. The session survives a harness failure
A new harness reads the log and resumes from the last event
14. After drops: all events or only items
Anthropic returns every event; OpenAI returns saved items
15. Nobody restarts a command after a disconnect
With your own environment, recovering compute is your job
16. Idle does not yet mean success
Success is confirmed by the turn outcome, tool results and the external effect
17. History returns; an unfinished action does not
Four kinds of stop and what each one leaves behind
18. Secrets are substituted on the way out
Agent code sees a placeholder; the real token is kept outside
19. A human-approval pause is a separate setting
Anthropic sets a policy per tool; OpenAI sets a mode per session
20. Your own sandbox brings operations back
Files and processes are yours; the loop and history stay with the vendor
21. Model review is not a security boundary
Network rules, separate environments and secret placement stay with you
22. These are vendor descriptions, not measurements
Five caveats to read the comparison with
23. Answer six questions before the handover
Two each for state, stops and control
24. Hand over the loop, keep the consequences
The platform keeps history; the sandbox keeps files
The session survives a harness failure
You verify the unfinished action
A human-approval pause is a separate setting
Your own sandbox brings operations back
Rehearse a stop before the first long task