Coding agents

• v2.1.276 fixes a 2.1.275 regression in Claude Code where every request failed with 400 … Input tag 'advisor_20260301' when ANTHROPIC_BASE_URL points at a proxy or gateway, so upgrade if you route through one.
• stablyai/orca is an ADE for desktop, mobile and remote that runs Codex, Claude Code, OpenCode or Pi side by side, each in its own worktree, with phone notifications and follow-ups when an agent finishes.
• Theo on Opus 5.5 filing a PR reports that a vague prompt produced a ready-to-merge T3 Code pull request with a video demo attached, an anecdote about agents owning the full PR loop.
• Theo on Claude Code vs Codex claims the $200 Claude Code plan is now far ahead of Codex, reversing the picture from a few weeks ago; sentiment only, no data.
• When chat is the wrong UI is GitHub's pitch for Copilot "canvases" as a more tangible interface than a chat box.
• AI-powered fuzzing with the GitHub Security Lab Taskflow Agent walks through a new fuzzing taskflow built on the Taskflow Agent framework, a concrete example of agent-driven security tooling.
• Default Enablement of Copilot Features for Copilot Business and Enterprise adds a global default policy enabling GA Copilot features and supported client capabilities in enterprise and org settings, with a 28-day window, so check admin settings now.

Models and benchmarks

• Terminal-Bench-Science 0.1 leaderboard launches an agentic benchmark for scientific research work, with GPT-6 Astra (max) at 63% and Claude Opus 5.5 (xhigh) at 62%.
• GLM-5.2 is the step change for open agents argues GLM-5.2 is a real jump for open-weight agents, and the post also announces a new paper on open RL recipes for terminal agents.
• Jev, a "System One Model" only decides, classifies, routes and scores, claiming >100x faster and >200x cheaper than small frontier LLMs, which matters for triage and routing layers in agent pipelines.
• Teaching Everyone to Fish for Tokens argues the Linux analogy for open models leaves a narrow path to a self-sustaining open-model ecosystem.
• 5 useful things in the post-training textbook announces the finished Manning book Reinforcement Learning from Human Feedback: Aligning and Post-training LLMs.

Agent memory and context

• vectorize-io/hindsight is an agent memory system aimed at agents that learn over time rather than just recall history, positioned as better than RAG and knowledge-graph approaches on long-term memory.
• AI Engineering bonus thread lays out the full production stack (context, agent loop, memory, tools, harness, guardrails, execution, verification, observability, evals, improvement) around the model; a useful checklist, nothing new.
• Solo founder on project context says proper project context helped more than any model upgrade; title only, no body captured.

Vibe-coding and repos

• mattpocock/skills is a set of small, composable agent skills that work with any model, pitched against process-owning frameworks like GSD, BMAD and Spec-Kit that take control away and make bugs hard to trace.
• Raft goes source-available and accepts "prompt requests" instead of pull requests, an experiment in an open-source contribution model for the agent era.

Agentic SaaS

• superdesigndev/treg is OpenRouter for agent tools: one base URL and token reach 3,000+ endpoints across 60+ providers, priced per call from a cent, plus your team's own keys, skills and CLIs.
• anthropics/financial-services provides reference agents, skills and data connectors for investment banking, equity research, private equity and wealth management, deployable as a Claude Cowork plugin or through the Claude Managed Agents API.
• Underwriting Superintelligence: Backing Agents you can Sue covers AIUC's $40M Series A and AIUC-1, an agent standard backed by real insurance, from a founder who was Anthropic's first product hire.
• Parallel cut research time and cost in half with GPT-6 Astra is a vendor case study of agents synthesizing labor-market data in half the time and at half the cost of prior models.