Scan window: 2026-09-13 16:00:31Z to 2026-09-14 16:02:27Z; retrieval overlapped back to 13:00:31Z. Added 9 events, updated 0 and skipped 1 duplicate release.

Daily Executive Summary

  • Agent tool interfaces produced an actionable pair of findings. Bash beat typed tools by 4.8-24.5 points across two enterprise-task benchmarks, while OATS showed why clean skills still need runtime consequence controls. Together with enterprise permissions, capability contracts and agent firewalls from the past week, they open trend #7: broad execution and consequence control are separating into layers.
  • A same-model paired study found no average success-rate advantage for the Claude Agent SDK or Codex SDK over a neutral harness. The clear difference was cost: the neutral harness spent 1.2-1.6 times more per solved task. ParaRecover separately turns parallel tool-error localization and replanning into process metrics.
  • AI infrastructure research focused on efficiency boundaries. AMD released large-scale ROCm-kernel training data, while SAS, SQD and a constant-state diffusion cache attack training, decode placement and cache complexity. The results are promising, but the latter three still lack public weights or independent reproduction.
  • An expert physics-benchmark audit traced 148 of 152 disputed cases to benchmark or grader errors. Frontier-model scores rose sharply after correction, making expert review a pre-release requirement for specialist benchmarks.

Models

  • No qualifying new model or production API release appeared in the window. SAS and the constant-state diffusion cache are notable architecture research, but neither provides directly adoptable model weights.

Agent & AI Engineering

  • Bash tool interface (TRIAL): It gained 21.8-24.5 points on TheAgentCompany and 4.8-7.4 points on APEX while using 19-72% fewer tokens. Any trial needs a strict sandbox and a separate policy layer.
  • Paired harness evaluation (TRIAL): Same-model, same-task results did not show an average success-rate advantage for vendor-native runtimes. Selection should include success, cache-corrected cost and observability.
  • OATS (TRIAL): It builds a consequence taxonomy from 66,192 skill versions. Its live gate blocked 23/23 prohibited attempts with 67.6 ms median hook latency.
  • ParaRecover (TRIAL): Its 10,626 examples cover 14 parallel-tool error classes and test whether an agent can locate the failing branch and rerun only necessary work.

Open Source

  • AMD released AMDKernelVault's HIP, Triton and ROCm-QA data plus Apache-2.0 code for training and evaluating kernel agents.
  • OATS and ParaRecover publish implementations. SAS publishes code but its repository currently declares no license, so production use needs an authorization check.

Research

  • SAS (WATCH): In continued training of OLMo3-7B, language-modeling loss directly trains token selection. Aggregate short-task results stay close to dense attention and long-context results are slightly higher.
  • SQD (WATCH): It splits decode around subquadratic attention stages. An 8×B200 proxy experiment and Rubin/LPX modeling report 31-56% better energy efficiency; end-to-end gains still depend on future hardware and system implementation.
  • Constant-state diffusion cache (WATCH): Experiments on 3B models keep state and latency from growing linearly at 256K context and improve very-long retrieval. No public code or checkpoint is available.
  • Expert physics-benchmark audit (TRIAL): After fixing questions and graders, GPT-5.6 rose from 47.3 to 78.7 on the audited HLE subset. Raw leaderboards should not be treated as reliable capability boundaries.

Developer Tools

  • Codex published only an undocumented 0.155.0-alpha.4, so it was excluded. Gemini CLI's September 14 nightly used the same source hash as the prior two days and was skipped as a repackaged duplicate. Claude Code's latest stable version remains 2.1.270.

Infrastructure

  • AMD's data is ready for small kernel-agent fine-tuning and offline evaluation. SQD, SAS and the constant-state cache remain WATCH items pending public checkpoints, real serving-stack benchmarks and independent reproduction.

Business & Policy

  • No qualifying event changed model supply, API cost, chip availability or the open-source landscape. OpenAI's Perplexity/Astra page was a customer story without enough technical detail and was excluded.

Trend Signals

  • New trend #7: Agent runtimes separate broad execution capability from consequence control (emerging / Medium). Five signals span product GA, controlled experiments and an open-source gate. The engineering move is to make command generation and consequence authorization independently testable, auditable and upgradable modules.
  • Trend #3 (established / High) adds the paired harness study and ParaRecover. Runtime selection and recovery capability now have reusable measurement protocols.
  • Other trends received no evidence strong enough to change status. The physics audit is an evaluation-quality signal, not a substitute for safety review or formal verification.

Tech Radar

  • New Bash tool-interface experiment (Agent & AI Engineering / TRIAL): enterprise-task success and token efficiency improve, but only with strong sandboxing and a separate policy layer
  • New paired harness evaluation (Developer Tools / TRIAL): native runtimes show no average success advantage, so cost and observability belong in selection criteria
  • New OATS (Agent Security / TRIAL): static cleanliness does not guarantee runtime safety; test a live consequence gate at small scale
  • New AMDKernelVault (Infrastructure / TRIAL): large execution-verified data creates a reproducible entry point for ROCm kernel agents
  • New SAS (Models / WATCH): end-to-end sparse-selector training has potential but needs public weights and independent long-context reproduction
  • New SQD (Infrastructure / WATCH): energy results are promising, while end-to-end conclusions still depend on modeling and future hardware
  • New ParaRecover (Agent Evaluation / TRIAL): provides fine-grained baselines for locating, recovering and replanning parallel tool failures
  • New expert physics-benchmark audit (Evaluation / TRIAL): specialist evaluations need expert review and grader audits before release
  • New constant-state diffusion cache (Models / WATCH): the long-context complexity advantage is strong; wait for code, weights and real-serving validation

Current radar: ADOPT 6 / TRIAL 65 / WATCH 82. See the site's Insights page for the complete current radar.

Worth Trying

  • Run Bash-versus-typed-tools A/B tests on an existing coding-agent suite with a fixed model and budget; record success, cache-corrected cost and prohibited actions.
  • Put OATS or an equivalent deterministic gate around one low-risk skill flow and measure false positives, latency and policy explainability.
  • Use ParaRecover's error taxonomy to check whether the current parallel-tool pipeline can recover only the failed branch.

Watch Items

  • Wait for public checkpoints, licenses and independent reproduction for SAS, SQD and the constant-state diffusion cache.
  • Watch for a second agent platform to publish a unified shell, MCP and browser consequence policy with adoption telemetry.

Sources