Projects › Lists › ProjectTasks › View item

ContentType: ContentType.ProjectTask   Item: agt-83

Investigate the best harness for small free-model workers: split jobs into sub-tasks, or run another harness (smolagents
2026-09-24T17:51:00.3135186+00:00
System Account (SPICE.Web)
AGT.83
AGT
Investigate the best harness for small free-model workers: split jobs into sub-tasks, or run another harness (smolagents, OpenHands, Goose, opencode) in the Docker Sandbox
(empty)
(empty)
8
Deferred

AGT.82 measured it: openai-fast added and read the datasheet with tool_add_file / tool_get_file, then did not carry the 40-field ResearchReport to a filing in 10 rounds. The task is parked Deferred on my-worker-69/Tasks. Compare: (a) decomposition - the workflow files one small task per section (electrical, mechanical, reviews, offers), each with its own publish; (b) a code-agent harness (smolagents: about 30% fewer steps than JSON tool calls); (c) an off-the-shelf agent in the sandbox kit. Keep what holds today: least privilege from the composition, finish/verify/unknown, public-only fetch. Score on filed-and-correct against the golden values (Wisdom: Filed is not correct), not on filed.

(empty)
(empty)
(empty)
(empty)
(empty)
(empty)

Attachments

No attachments.

+ Attach a file
Status:Deferred
✓ This item has reached its final state.