Source: github.com/agno-agi/agno/releases/tag/v3.1.12
Native approvals for harness agents let Claude's and Codex's permission requests reach the AgentOS approval UI when the agent has a database. A native request persists a paused run that appears in the existing chat and Approvals views, and Resolve plus Continue Run sends the decision back to the same execution without replaying the prompt. Sandboxed agents on Docker and AgentCore keep this behavior. A CodexAgent with no approval settings now runs under the on-request policy, so requests pause the run, with hitl_timeout defaulting to 900 seconds.
Routine(enabled=False) lets a template ship an inactive Company Routine that a user can turn on later from the UI. Disabled Routines stay linked to their Product, are skipped by the scheduler, and keep the saved choice across restarts. GET /company/routines can now also sort by next run, schedule type and enabled state.
The AccuracyEval judge default has changed from o4-mini, which OpenAI shut down on 2026-10-23, to gpt-5.6-terra. Per-upload chunker options no longer persist on the shared knowledge base reader, and POST /company/routines/{id}/cancel now records a status only once the run has actually stopped.
Why it matters if you build agents
If you run Claude or Codex harnesses in AgentOS, permission prompts now pause runs for approval in the UI instead of going nowhere, so check your approval settings and hitl_timeout.
If you use AccuracyEval without an explicit judge model, the default no longer points at a retired OpenAI model, and evals that returned empty results on that default should be rerun.
Get stories like this every weekday
One short email with the day's agent tooling news.