Skip to content

AgentOS 2026.7.19

Choose a tag to compare

View all tags
@andreapn andreapn released this 19 Jul 13:06
v2026.7.19
d4c7dcc
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

🚀 Headline — default router change (the biggest highlight)

PR #26 — AgentOS Pilot (pilot-v1) — an English-language, self-trained, on-device router that replaces v4_phase3 as the default routing strategy.

  • ONNX classifier (~500 KB) owned end-to-end by AgentOS (replacing third-party weights that couldn't be retrained/debugged), with an all-MiniLM-L6-v2 INT8 backbone (Apache-2.0).
  • Benchmark (983-turn sealed set): accuracy 0.260 → 0.626 (+36.5pp), under-routing 0.646 → 0.152, R2 recall 0.066 → 0.698. Wins on 11/12 axes; the trade-off is a slight increase in over-routing (0.094 → 0.223).
  • v4_phase3 is still selectable — rollback is a 1-line config change.
  • Includes a reproducible training pipeline in scripts/pilot_router/.

PR #36 — force-migrate v4_phase3 → pilot-v1 on config load ⚠️ (worth calling out explicitly)

  • Existing configs that persist strategy = "v4_phase3" are automatically rewritten to pilot-v1 (with a timestamped backup, idempotent) on the first agentos command after upgrade. No user action required.
  • The selector CLI and web setup drop the v4 option (now 3-way: Pilot / LLM judge / Off).

PR #34 — remove the T1 benchmark report from the public tree.

✨ New Features

PR #38 — agentos upgrade (single upgrade command)

  • Upgrade + restart the managed gateway + verify, automatically.
  • Includes: version-compare, a PyPI client, install-method detection (uv-tool, custom UV_TOOL_DIR), a version-skew policy, a passive update notice, and a new updates.notify config key.

PR #30 — local providers work end-to-end (Ollama / LM Studio / OVMS / vLLM)

  • Degrade to a single model for local providers, surface router-tier ↔️ provider mismatches at boot and in doctor, and keep the local tier pinned when the router is disabled.

PR #37 — bundled agentos self-operation skill (merged).

PR #39 — Bankr skill polish — limit to 2 skills, an Update button, emoji avatars, brand-glyph logo fallback, and filled-in browse descriptions.

🧹 Housekeeping

PR #40 — bump version to 2026.7.19.

PR #41 — rename the display name "AgentOS Router" → "Pilot Router" (display strings only; config keys, class names, and module paths unchanged).

Assets 6
Loading