Tessary runs cheap classifiers over every agent trace, then sends only flagged ones to SOTA-model agents for root-cause triage; it ingests OTLP with the gen_ai spec and can be self-hosted. The team says it came out about 5x cheaper than sampled evals at 1M traces.
Muse Code installs as a native muse CLI that plans, edits, and runs commands with approvals and an OS sandbox on by default, plus a headless muse exec mode for CI. It ships with skills, hooks, MCP support, and a TypeScript SDK, defaulting to muse spark 1.2…
Anthropic now lets a Managed Agents agent write a workflow that runs many agents in phases and combines their output as a background run. Turn it on by setting the agent's multiagent field to multiagent 20261001 and track progress via workflow run. events…
Pi 1.1.0 lets terminals and agent dashboards see whether it's working, blocked on a login, done, or failed via OSC 7501, and pi t +codemode, write now tweaks the default tool set instead of replacing it. It also adds Claude Haiku 5.5 with adaptive thinking and runs GPT 6 Luna and llama.cpp decision models as classifiers…
Copilot's local sandboxing, which limits what agents can touch on files, network, and credentials, is now generally available at no extra cost. The Copilot app also now lets you use one GitHub account for your license and another for your repos…
Vega pairs a frozen Qwen3.5 0.8B reader with a small trained engine to return choices, 0–1 scores, or true/false probabilities with an explicit abstain flag, plus 73k token context and image input. It's Apache 2.0 with a 4B variant, installable via pip install vegaml…
Spotted something we missed? Start a thread.