Scan window: 2026-09-22 16:02:37Z to 2026-09-24 16:01:46Z; retrieval overlap starts 2026-09-22 13:02:37Z. Nine new events, four existing events updated, and 11 duplicate, prerelease or low-value candidates skipped. Qwen Code 0.24.4 is an overlap recovery. The report date follows Asia/Shanghai.

Daily Executive Summary

  • GPT-6 Sol/Luna and Claude Opus 5.5 push model selection further from “which is strongest?” toward “what does a completed task cost?” Both target long-context, tool-using agent workloads, but at very different price points. Route them with your own completion rate, step count and cache-hit data.
  • Cursor extends coding agents past merge. Rollouts watches production telemetry and can prepare a pause or revert, while Security Review inspects pull requests with whole-codebase context. Approval and reversibility must expand with that authority.
  • GitHub, Qwen Code and PASTABench point to the same conclusion from product, runtime and evaluation layers: broad execution needs separate sandboxing, attestation, telemetry and timely intervention. Trend #7 moves from emerging to strengthening, still at Medium confidence.

Models

  • GPT-6 Sol targets complex coding and agents; Luna targets high-volume work. They share a 1.05M-token context and built-in tools, but their standard output prices differ by 20x, making them candidates for difficulty-based routing within one harness.
  • Claude Opus 5.5 lowers input/output pricing to $4/$20 per million tokens and cache reads to $0.20. METR's predeployment evaluation describes its AI-R&D gain as incremental, not a capability discontinuity.

Agent & AI Engineering

  • Cursor's production tests suggest an optimization order that other teams can reuse: shrink the stable prefix, load tool schemas lazily, stabilize cache boundaries, compress source-line context, then use subagents selectively.
  • PASTABench finds that even the best safety monitor intervenes inside the optimal window only 40.74% of the time. A safety classifier must be right before the harmful state transition, not only right afterward.

Open Source

  • Qwen Code 0.24.4 adds managed runtime attestation, durable tool-execution records and plan-drift checks, and moves bubblewrap isolation to tool execution. It landed 47 minutes before the prior checkpoint and is an explicit overlap recovery.

Research

  • About 950 Claude agents searched more than 200,000 reverse transcriptases in 21 hours, and a human-selected candidate reached wet-lab testing. The biological function is unresolved; this demonstrates an agent-hypothesis plus human-validation workflow, not autonomous science.
  • SWE-Flux derives 480 repository-level runtime questions from instrumented tests. The best model reaches only 37% accuracy, showing that editing code is not the same as understanding inter-procedural dataflow and state.
  • UK AISI publishes configuration-rich, machine-readable evaluation records through EvalEval, adding government-evaluator adoption evidence for reproducible disclosure.

Developer Tools

  • Cursor Rollouts creates a monitoring plan from a PR diff and can notify, pause a progressive deployment or prepare a revert PR under policy. Start in notify-only mode and measure false positives before widening authority.
  • GitHub Copilot app adds enterprise-managed OTel and local sandboxing with fail-closed enforcement. JetBrains 1.18 adds assisted approvals and persistent per-MCP-tool controls.
  • Codex 0.156 makes worktree sessions a default command-center capability and adds sandbox fixes, usage analytics and GPT-6 Sol/Luna routing. Claude Code 2.1.280/281 adds Opus 5.5, Bedrock role/guardrail controls and stricter auto-mode checks.
  • Gemini CLI 0.61 promotes build-file indirect-prompt-injection defenses and filesystem isolation to stable; its A2A server remains undocumented.

Infrastructure

  • No new training or inference infrastructure event cleared the threshold. Continue watching the reported vLLM 0.30 CPU-wheel compatibility issue for an official fix; an unresolved user issue is not stored as an event.

Business & Policy

  • UK AISI's EvalEval adoption matters because it records configuration, inference-time compute and method alongside the score. Cross-model claims become genuinely comparable only if more labs reuse the same schema.

Trend Signals

  • Trend #3 gains production evidence: Cursor enters the deployment feedback loop, Codex defaults to worktree sessions, and Qwen Code adds runtime attestation. Coding agents keep turning from chat tools into orchestrated runtimes.
  • Trend #4 gains complementary evidence: METR publishes both its Opus 5.5 predeployment findings and its independence terms, while UK AISI publishes machine-readable results through EvalEval. A cross-vendor disclosure format is still missing, so it stays strengthening / Medium.
  • Trend #7 gains cross-organization product and benchmark evidence: GitHub fails closed when local sandbox policy cannot be enforced, Qwen Code adds attestation, and PASTABench measures intervention timing. It moves from emerging to strengthening / Medium.
  • No new trend has enough evidence in this cycle.

Tech Radar

  • New GPT-6 Sol/Luna (foundation-model / TRIAL): shared long-context and tool surface with sharply different costs; test difficulty routing on real work.
  • New Claude Opus 5.5 (foundation-model / TRIAL): cache and task costs fall materially, but vendor benchmarks still need harness-specific validation.
  • New Qwen Code 0.24.4 (developer-tools / TRIAL): attestation, execution records and plan drift enter stable for managed-agent trust trials.
  • New UK AISI EvalEval (business-policy / WATCH): machine-readable evaluation disclosure gains a government adopter; wait for broader reuse.
  • New Cursor Rollouts and Security Review (developer-tools / TRIAL): agents enter the deployment loop; begin with notification and suggested fixes.
  • New Cursor context optimization (ai-engineering / TRIAL): production evidence gives a low-risk token-reduction path, with task regression after each change.
  • New Claude biology workflow (research / WATCH): the hypothesis reached wet-lab tests, but its function is unresolved and does not prove autonomous science.
  • New PASTABench (agent-security / TRIAL): intervention timing becomes measurable for long-horizon safety acceptance.
  • New SWE-Flux (research / TRIAL): execution traces provide non-LLM-judge ground truth for repository-level debugging evaluations.
  • Codex CLI (developer-tools / ADOPT): 0.156 ships worktrees, usage analytics, sandbox hardening and GPT-6 Sol/Luna routing.
  • Claude Code (developer-tools / ADOPT): 2.1.280/281 strengthen symlink writes, Bedrock guardrails and auto-mode review.
  • Gemini CLI A2A (developer-tools / WATCH): upgrade for 0.61 security fixes, but the A2A server remains undocumented.
  • GitHub Copilot managed permissions (developer-tools / TRIAL): local sandboxing and OTel add prevention and investigation, but require explicit configuration.
  • Current radar: ADOPT 11 / TRIAL 97 / WATCH 100. See the site Insights page for the full current radar.

Worth Trying

  • Run the same coding-agent tasks on GPT-6 Sol, Luna and Opus 5.5; record completion, steps, cache reads, total latency and cost per completed task.
  • Enable Cursor Rollouts in notify-only mode on one low-risk repository and compare its monitoring plan with current SLOs and release controls.
  • Add PASTABench-style Earliest-Signal and Trigger labels to one long-running agent workflow and test whether the monitor can stop the state transition.
  • Start context budgeting with the system prompt and tool schemas. Run regression after each change instead of changing cache, tool loading and subagent topology together.

Sources