CEREBRO

Coding agents & models

Introducing Claude Opus 5 Anthropic's new flagship, pitched as "thoughtful and precise"; Simon Willison hasn't stress-tested it yet but early buzz is positive.
Claude Code v2.1.219 ships Opus 5 as default Opus model (1M context, fast mode $10/$50 per Mtok), adds sandbox.network.strictAllowlist to hard-deny non-allowlisted hosts and a DirectoryAdded hook.
Opus 5 now in GitHub Copilot targets complex long-running coding tasks needing sustained reasoning + tool use.
Opus 5 #1 on Artificial Analysis Intelligence Leaderboard but community pushback: real-world agentic performance reportedly lags Opus 4.8/Fable, poor cost-to-intelligence ratio vs rivals at half the price.
GLM-5.2 is the step change for open agents Nathan Lambert flags a capability threshold crossed for open-weight agentic use, backed by new open RL recipes for terminal agents.
Open model bonanza roundup Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, GLM-5.1 all landed this month; CAISI ran a fresh V4 safety/capability assessment.

jcode new agent harness pitched on multi-session workflows, "infinite customizability," and a raised skill ceiling, one more entrant vs Claude Code/OpenCode.
The most annoying Claude Code footgun: mismatched effort levels reasoning-effort setting silently mismatched to task cost is a recurring complaint worth checking your own config against.
The one habit that made Claude Code trustworthy on big multi-file changes discipline-based workaround thread for multi-file reliability, worth a skim for process tips.
Chimlo tracks Codex/Claude Code sessions and lets you respond from the MacBook notch, niche but on-trend for ambient agent monitoring.

Orca ADE for running Codex/Claude Code/OpenCode/Pi side-by-side in separate worktrees under one pane, plus quick-open search and unified usage tracking across accounts.
Why AI infra must evolve for Agent Experience Modal CTO argues cloud infra was built for humans, not agents, context: Modal just closed a $355M Series C.
OpenAI GPT-Live desktop app for agent orchestration OpenAI pushing an infra/orchestration play this week, competitive pressure on the agent-fleet-management space.
The great loops debate — AIEWF Daily Dispatch AI Engineer World's Fair closed on a debate over agent loop design, relevant framing for anyone building agentic control flow.

mattpocock/skills real-world Claude Code skills from an engineer's actual .agents directory, positioned against heavier frameworks (GSD, BMAD, Spec-Kit) that take over your process.
OmniRoute free-tier LLM gateway aggregating 43 provider pools/500+ models behind one endpoint, quota-aware fallback and claimed 15-95% token savings via compression, works with Claude Code/Codex/Cursor.
Automating cross-repo docs with GitHub Agentic Workflows turns merged product changes into SME-reviewed doc PRs automatically.
How Codex became a collaborator for OpenAI's creative team internal case study of Codex used outside pure engineering workflows.

SnapState persistent state layer for AI agent workflows, one more entrant in the agent-memory/checkpointing space.
HarnessRouter one API to plug top AI agents into your app.
Firecrawl's new /search accuracy-focused search API built specifically for agent consumption.
MCP storefront for Claude shopping full store exposed as MCP, Claude can search/design/cart/checkout end to end, concrete example of commerce-via-agent.
MCP server for unused Veo video credits small but illustrative pattern: MCP wrapper built just to stop wasting a subscription's included quota.
Claude given keys to a UniFi network home-network agentic control demo, useful signal on real-world tool-use scope creep.