Microsoft Agent Framework’s Sep. 18 releases add production plumbing across the stack: Python 1.19.0 introduces generic vector-store protocols plus MongoDB, DocumentDB, and Cosmos connectors, stable orchestration names/checkpoint restoration, Dev UI Aspire traces, sequential tool invocation, and tighter security labels; .NET 1.22.0 is released alongside it. The notable direction is convergence of memory, observability, checkpointing, and policy controls in the framework rather than another agent-demo feature. Official release: https://github.com/microsoft/agent-framework/releases
GitHub Copilot CLI 1.0.87 (Sep. 21) adds managed and user startup defaults for the Auto routing tier, strict/overridable organization policy, recallable queued steering prompts, and configurable worktree path templates. This is a meaningful harness update because it moves routing policy and long-running local-session ergonomics into the CLI itself; the release is official, with community discussion around the broader Copilot CLI rollout still active. Sources: https://github.com/github/copilot-cli/releases ; https://www.reddit.com/r/GithubCopilot/comments/1wmfcx7/github_copilot_for_jetbrains_v118_updates/
A new research direction is treating the harness—not only the model—as the unit of continual improvement. Harness-of-Harness (arXiv 2609.01481) iterates planning, coding, testing, and repair over existing coding-agent harnesses and reports average relative gains of 52.25% after three iterations across GameCraft-Bench, FrontierSWE, and ProgramBench; the claims are paper-reported and need replication. Source: https://arxiv.org/abs/2609.01481 ; project: https://flesymeb.github.io/HarnessOfHarness/
FrogNano is a newly surfaced small-model coding-agent result: a 4B model trained with online task synthesis at the current model’s learnability frontier, rather than relying on distillation. It is relevant to local/cheap agent deployment because the paper explicitly targets competitive software-engineering performance at small scale; benchmark claims are research-reported, not yet independently established. Sources: https://arxiv.org/abs/2609.07925 ; https://huggingface.co/papers/2609.07925
RSIAgent explores frozen-weight recursive improvement through autonomous environment exploration: the agent chooses experiments, executes and verifies tasks, and stores reusable action-condition-outcome memories. The paper reports gains on OSWorld-v2 and Agent’s Last Exam, including claims that Kimi-K3 and GLM-5.3 can exceed frontier closed models under this scaffold; treat those results as preliminary until independently reproduced. Sources: https://arxiv.org/abs/2609.15364 ; https://github.com/AetherLabsAI/RSIAgent
An unfamiliar model-architecture lead from the discovery pass is “The Router Within”/Gavel, which embeds skill routing inside a frozen language model and combines a fast glance with the model’s own yes/no skill judgment. The authors report zero-shot transfer on three public benchmarks and a 372-trajectory SkillTraj benchmark; this is an early research signal for structured decision models that can reduce unnecessary tool/skill execution, not a production result. Source: https://arxiv.org/abs/2609.15982 ; discovery index: https://arxivsignals.io/papers/2609.15982