Topic
Publicly released software—agent frameworks, harnesses, and developer tools—shared by AI labs and chipmakers to let others build, run, or optimize AI agents.
Microsoft Research Asia open-sourced Agent Lightning v1.0, a 3,500-line framework whose recipe lifted a 9B model's SWE-bench Verified score by 14.6 points. It runs reinforcement learning on the agent a team already deploys, sitting as a proxy between that agent and its model.
Reality
Google Cloud AI Research has released RRSI, an Apache 2.0 tool whose self-rewriting agent harness lifted Terminal-Bench 2.1 scores from 74.2% to 80.2%. Any team can use it commercially, though on tasks the agent never trained against the reported gain falls to 4.7 points.
Reality
Intel has added NVIDIA's Apache-2.0 OpenShell policy layer to its enterprise agent toolkit, putting sandboxing and kernel-level agent controls on Xeon. The feature ships switched off, so what it is worth to Intel depends on how many agent teams turn it on.
Reality
Google open-sourced AX, an Apache 2.0 runtime on Kubernetes that checkpoints idle agents and is designed to resume them in under a second. Its savings depend on how long each agent waits on models, tools or people.
Reality
Harness v0.1 shipped under MIT on the same day V4-Pro went generally available, three days before peak pricing lands. The lock-in it targets is the runtime, not the weights.
Perspective Coverage
4 publishersReality
AWS has open-sourced a preconfigured general-purpose agent on top of its Strands SDK and says it runs 45% cheaper than Claude Code and Codex. The model call defaults to Amazon Bedrock, and Marc Brooker says one line changes that.
Reality
Anthropic's open-source audit framework now runs a classifier over every auditor turn and rewrites anything a real deployment would not produce. The tuning targeted models that say out loud they are being tested.
Reality
Ao Qu and collaborators open-sourced Reef on September 15th, an OpenAI-format inference server that stamps each response with a record ID, matches later feedback to that ID, and publishes a retrained artifact only if its evaluation stage accepts it.
Reality
CauterRule's own field test puts the undecided bucket above pass and fail combined, and the report names the matcher that produced it as its top calibration target. The per-model rates end up measuring trigger phrasing.
Reality
Nvidia's Apache 2.0 beta seizes the port Ollama or LM Studio was using and forwards each call to whichever machine is free, which makes local throughput a function of how many PCs you own. Hands-on testing found it serving an engine Nvidia does not list.
Reality
Headlong is Apache-2.0, so a team can read the loop before running it. What nobody can read is a benchmark, and the shared thought stream is also the privacy problem.
Reality
DeepSeek Harness is an MIT-licensed developer preview in which the model adapter, tool registry, session log and agent loop are all plugins. The interesting part is what that implies for evaluation.
Reality
The best of 12 models passed 65.36% of 507 business tasks on the first attempt and only 25.25% across all 20 trials. Agent procurement should be priced on the second number.
Reality
An open-source stack pairs a deterministic Minecraft reimplementation with seed-level provenance, so a reinforcement learning result can be replayed instead of reconstructed.
Reality