Today: 1 new event and 2 updated events. Scan window: 2026-09-27T16:01:12Z to 2026-09-28T16:03:45Z with a three-hour retrieval overlap; six seven-day reviews completed.

Daily Executive Summary

  • H Company released Holo4, an open-weight agent family that uses one harness across GUI actions, code and MCP/API tools. Weights, harness and trajectories ship together, but the 27B model is non-commercial and vendor benchmarks include single-run and training-split caveats; start at TRIAL.
  • Codex 0.158.0 stable adds client secrets for pre-registered MCP OAuth clients and bearer-token authentication for exec-server, and enables terminal-input approval by default for elevated commands. It also fixes sandbox boundaries across Windows, Linux and macOS. Keep ADOPT.
  • All six research reviews due today were completed. Public artifacts for DUMA-Bench, RRSI and MSI-Bench remain available, but no independent reproduction appeared; the other three papers also gained no code or reproduction evidence. Ratings stay unchanged.
  • LiteLLM 1.103.0 stable adds a Claude Code gateway, RFC 8693 token exchange, MCP client allowlists and session revocation. The release spans a large RC range; keep TRIAL and run production-shaped regressions.

Models

  • No new general-purpose foundation model shipped. Holo4 is a computer-use agent model and is covered under Agent & AI Engineering.

Agent & AI Engineering

  • Holo4 combines GUI, code, MCP and API work in one agent and harness. It is an inspectable engineering release, but the vendor's single OSWorld 2.0 run is not independent validation.
  • Codex 0.158.0 makes enterprise MCP servers that require pre-registered credentials easier to connect through OAuth client secrets. This is one stable client capability, not evidence that cross-vendor MCP authentication has converged.
  • Terminal-input approval is now on by default for elevated commands. For long-running work, the important change is an independent consequence check on another side-effecting runtime channel, not merely another dialog.

Open Source

  • Holo4-35B-A3B is Apache 2.0 and can enter a commercial POC; Holo4-27B is CC BY-NC 4.0 and should remain research-only without separate permission. Both provide ungated weights.
  • At the seven-day review, RRSI had 459 stars and 47 forks, with no push after September 23. Adoption is promising, but no independent team has reproduced its held-out benchmark gains; keep WATCH.

Research

  • Seven-day reviews completed for OSWorld-Pro, DUMA-Bench, RRSI, trajectory-sampled regression, emergent agent collusion and MSI-Bench. Public materials remain available, but none gained independent reproduction that changes maturity or recommendation.
  • The fixed arXiv feed returned submissions no newer than September 25, before this window; no new cs.AI paper was stored.

Developer Tools

  • Codex 0.158.0 is one of two stable tool releases that cleared the update threshold. Beyond MCP and exec-server authentication, it fixes ordinary Windows 10 paths, credential repair, large permission policies, nested Linux writable roots, Git metadata protection and macOS path aliases.
  • Gemini CLI 0.63 remains nightly-only and Codex 0.159 remains alpha. Neither is treated as a stable release.

Infrastructure

  • LiteLLM 1.103.0 strengthens gateway identity, MCP-session and budget controls. Its release notes aggregate many RC changes, backports and reversions, so configuration-level regression testing remains necessary.
  • vLLM 0.30.0 remains the latest stable release, with no newer tagged fix.

Business & Policy

  • No in-window event materially changed model access, API cost, chip supply or the open-source landscape.

Trend Signals

  • Trend #2 gains two stable-release signals from Codex and LiteLLM: MCP client secrets, token exchange, client allowlists and session revocation enter products. Cross-vendor authentication and enforcement schemas remain absent; stays emerging / Medium.
  • Trend #7 gains two runtime-policy signals: Codex defaults elevated commands to input approval, while LiteLLM keeps MCP-aware guardrails and fail-closed credentials in the gateway. A unified policy schema remains absent; stays strengthening / Medium.
  • The other five active trends received no new independent signal that changed their assessment.
  • No evidence-supported new trend emerged in this cycle.

Tech Radar

  • New Holo4 (ai-agent / TRIAL): weights, harness and trajectories ship together, but verify licenses and rerun benchmarks independently.
  • LiteLLM (ai-engineering / TRIAL): 1.103.0 strengthens gateway identity, sessions and budgets; regress production configuration.
  • Codex (developer-tools / ADOPT): 0.158.0 moves MCP credentials, exec-server authentication and elevated-command approval into stable while fixing several cross-platform sandbox boundaries; rating unchanged.
  • Current radar: ADOPT 11 / TRIAL 107 / WATCH 104. See the site insights page for the complete current radar.

Worth Trying

  • For a commercial trial, start with Apache-2.0 Holo4-35B-A3B and rerun OSWorld 2.0 plus the 120 held-out AutomationBench tasks in your own harness.
  • After upgrading Codex 0.158.0 in a test environment, regress three boundaries: pre-registered MCP OAuth client secrets, bearer-token exec-server WebSockets and elevated commands that need terminal input. Run Windows and nested writable-root cases separately.

Sources