AI AGENTS
446 items
440 posts, 2 tools, 4 guides
Anthropic is cutting its vulnerability disclosure window from 90 days to 30 for high and critical bugs, effective January 1, 2027. What changes, what does not, and a Monday checklist for maintainers.
FrontierHarness ran one model through 9 coding agent harnesses. Pass rates sat within 17 points, median cost per pass spread 17.5x. What to measure before you pick Claude Code, Codex, Pi or OpenCode.
Claude Code Code Mode lets Claude write JavaScript that makes several tool calls in one step. Turn it on with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 on v2.1.323+.
Self-hosted Claude Managed Agents runs the service, its PostgreSQL store and sandboxes on your infra, with Bedrock, Vertex, Foundry or a gateway for model calls.
A viral essay argues AI-written code makes codebases fragile: shallow tests, needless complexity, lost understanding, no time to review. What holds up, what is anecdote, and four checks to run on your own repo.
Google DeepMind used AI agents to prove security properties of the libsignal machine code in Lean. What the paper actually claims, what you must still trust, and what a builder can copy for agent-written code.
Amazon engineers replaced Sonnet 4.6 on a 20-label ticket classifier with Strands Decider, a small decision model, at about 44x lower list cost. Here is the narrow job it fits and the rollout pattern worth copying.
Claude Code 2.1.319 adds permissions.denyInsteadOfPrompting: any call that would wait on a permission prompt is denied and the turn carries on, subagents included.
OpenAI built Pando, a Git-compatible backend, after Codex flooded its monorepo with pushes. What it changed, the trade-offs, and what it means for your agent fleet.
Claude API message threads (beta, Nov 2) keep a conversation on the server for 5 minutes, so each turn sends only new messages. Setup, SDK helper, errors, limits.
Since Claude Code 2.1.314, one setting makes Write, Edit and NotebookEdit refuse paths outside your working directories in every mode, including bypassPermissions.
Oh My Pi (omp) is a batteries-included fork of the Pi coding agent. What it adds, how its model roles and fallback chains work, and where to be careful.
T3 Code is a free, MIT-licensed desktop, web and mobile app that drives Claude Code, Codex, Cursor, OpenCode and more. What it is, how remote access works, and what to check first.
DeepSeek open-sourced dsh-superagent, an always-on coding agent built on a router and worker split. What the design says, what the code ships, and the safety debate.
Claude Code supports MCP Tasks and elicitations. What a tool call needs to become a task, what v2.1.319 changed for Bedrock and gateways, and what server authors change.
Langfuse is part of ClickHouse, Promptfoo is part of OpenAI, Galileo is part of Cisco, Traceloop is joining ServiceNow, Helicone joined Mintlify and Humanloop shut down. Who owns what in October 2026, the dates the vendors state, and a portability checklist built on OTLP ingest, licenses and self-hosting.
MCP sampling, elicitation and roots explained against the 2026-07-28 spec, with JSON-RPC examples and which of 24 clients support each one.
GPT-Live can now call the Decisions API itself, mid-sentence. How native delegation and transcript triggers work, the session config, and the events to handle.
Agent memory is five layers: history, compaction, semantic recall, working memory and files. How Mastra, LangGraph, OpenAI and Claude implement each.

Get Smarter About AI Dev
New tutorials, open-source projects, and deep dives on coding agents - delivered weekly.