Scan window: 2026-09-15 16:00:31Z to 2026-09-16 16:00:47Z; retrieval overlap starts 13:00:31Z. Eight new events, two existing events updated, and four prerelease or low-value candidates skipped.

Daily Executive Summary

  • This cycle is about verifying training, evaluation and agent orchestration rather than a new model launch. OPEN-1B makes individual training steps bitwise replayable across hardware. A SWE-bench audit finds that small gaps among leading coding agents do not provide statistical ranking resolution.
  • Two multi-agent results complement each other. Production traces show that deeper hierarchies lose information at handoffs, while a cross-principal experiment finds that current messaging primitives fail as more honest agents participate. Gemini CLI 0.60 and Claude Code 2.1.273 also harden permission, provenance and runtime boundaries.

Models

  • OPEN-1B (TRIAL): The release includes a 400B-token training trajectory, 810 checkpoints and an audit-replay harness. Each step can reproduce the same bits on CPU, CUDA or Metal, making it useful for model-provenance experiments.
  • No new frontier foundation model or GA API release met the bar.

Agent & AI Engineering

  • Multi-agent decomposition study (TRIAL): Across 600 production research traces and 743,819 tool calls, deeper delegation reduces root-context pressure but loses findings and increases drift. Under the paper's measured parameters, two tiers beat a flat agent at equal cost only beyond about 403 findings.
  • ScienceBuddy (TRIAL): User feedback and execution evidence become evaluation rubrics; the system improves the harness before training the model. Code and a workspace are public, but held-out evaluation and human review remain necessary.
  • Social harness (TRIAL): In cross-principal scheduling, one configuration falls from 90% success with one participant to 0% with seven. A2A supplies connectivity; provenance, authority, task protocols and consequences need an independent layer.

Open Source

  • OPEN-1B, ScienceBuddy, the social-harness experiments and the SWE-bench resolution audit release code or reproducibility artifacts. JustFit provides an evolving MLX branch, though exact historical environments are incomplete.

Research

  • SWE-bench resolution audit (TRIAL): None of the 29 adjacent pairs in the Verified top 30 is separable by an exact paired test. Within-model scaffold ranges reach 29.8 points. Selection should use internal tasks, paired uncertainty and the model-scaffold combination.
  • These findings extend the evaluation side of trend #3: coding-agent outcomes increasingly depend on the runtime and scaffold. They do not change the established / High adoption judgment.

Developer Tools

  • Gemini CLI 0.60 (material update): Stable now enforces RFC 9207 issuer identification in MCP OAuth and hardens WebFetch, extension paths, environment changes, tool-output provenance and sandbox directories.
  • Claude Code 2.1.273 (material update): Adds gateway routing headers, MCP reconnect-failure notices and Remote Control session forks. It also fixes skipped Bash prompts, hidden destructive subshells, managed-policy precedence and external-memory reads.
  • GitHub AI Scan (TRIAL): Pull-request scanning no longer requires CodeQL default setup on each repository. It remains a GitHub Advanced Security public preview and does not support GHES.

Infrastructure

  • JustFit (TRIAL): Qwen3.8-27B MXFP4 completes 196K input plus 16K output on a 24 GiB M4 Pro. A limit run takes about 103 minutes, making asynchronous ingestion and prefix reuse the practical use case.
  • FlashVector (WATCH): Unity reports production serving optimization across CUDA, PyTorch, Triton and Python feature layers, with model-server throughput up to 2x. The implementation is closed and awaits independent reproduction.

Business & Policy

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

Trend Signals

  • Trend #3 (established / High): The SWE-bench audit and decomposition-cost study make runtime, scaffold and delegation topology more measurable. They improve engineering decisions without changing adoption maturity.
  • Trend #7 (emerging / Medium): Gemini CLI 0.60, Claude Code 2.1.273 and the social-harness experiments support separating capability, communication and consequence control. A second platform with one GA policy across shell, MCP and browser plus adoption telemetry is still missing.
  • No new trend has enough independent evidence in this cycle.

Tech Radar

  • New OPEN-1B (Open Source / TRIAL): step-level cross-hardware replay adds a verifiable training layer to open models
  • New SWE-bench resolution audit (Agent Evaluation / TRIAL): frontier score gaps do not support strict ranks; use paired tests and tiers
  • New multi-agent decomposition rule (Agent & AI Engineering / TRIAL): deeper delegation loses findings; measure the local cost crossover first
  • New JustFit (Infrastructure / TRIAL): consumer hardware can hold 200K context, but limit latency is high
  • New ScienceBuddy (Research Agent / TRIAL): trial the trace-to-rubric loop before recursive model training
  • New social harness (Agent Security / TRIAL): cross-principal messaging needs provenance, authority and protocol outside the model
  • New FlashVector (Infrastructure / WATCH): cross-stack production gains are promising, but code and independent reproduction are missing
  • New GitHub AI Scan expansion (Developer Tools / TRIAL): broader coverage warrants a preview trial with precision and review-load measurement
  • Gemini CLI (Developer Tools / WATCH): 0.60 moves security fixes to stable; the A2A server remains undocumented
  • Claude Code (Developer Tools / ADOPT): 2.1.273 closes more permission and managed-policy gaps; upgrade is recommended
  • Current radar: ADOPT 6 / TRIAL 74 / WATCH 85. The site insights page holds the full current radar.

Worth Trying

  • Run per-instance paired tests on existing coding-agent results before trusting leaderboard gaps.
  • Record finding retention at each research-agent handoff and compare token cost for flat and two-tier designs.
  • Replay one OPEN-1B step in isolation to estimate the operational cost of verifiable training governance.

Watch Items

  • Await independent reproduction of FlashVector, the decomposition crossover and social-harness results.
  • Watch for Gemini CLI, Claude Code or a second platform to publish one consequence policy across shell, MCP and browser with adoption telemetry.

Sources