CEREBRO

Coding agents & orchestration

Orca — ADE running Codex/ClaudeCode/OpenCode/Pi side-by-side, each in own worktree, one dashboard; own-subscription model, desktop+mobile.
Claude Fable 5 and new AI safety fables — Anthropic swapped silent prompt-manipulation for a classifier on safety queries after community pushback; matters for anyone routing sensitive research through Claude.
About the Safeguards in Fable 5 — community thread dissecting Fable 5 safety behavior, companion to above.
anthropics/claude-cookbooks — official copy-paste recipes for Claude API integration patterns.
mattpocock/skills — real-world .agents skill set built for actual engineering, positioned against process-heavy frameworks like GSD/BMAD/Spec-Kit that remove control.
unicity-aos/aos-ce — open, inspectable "agent operating system" (CLI, HTTP API, capsules) for composable agent environments.
AOS-style: Openbase — voice-driven management of a fleet of AI agents, ProductHunt launch.

Can you optimize for cache hits — practitioner thread on structuring prompts for Claude prompt-cache hit rate, directly relevant to token-cost tuning.
GLM-5.2 is the step change for open agents — open-weight model claims a capability threshold for terminal/agentic RL use, worth benchmarking against Claude/GPT for agent tasks.
[AINews] GPT 5.6 Sol/Terra/Luna, Codex becomes ChatGPT superapp — OpenAI folds Codex into ChatGPT as a superapp surface, competitive pressure on Claude Code's agentic-IDE niche.
Unpopular opinion: I really like working with Opus 5 — contrarian take on Opus 5 coding quality vs prevailing sentiment.
I feel like most people underestimate Sonnet for coding — pushback on Opus-only bias, relevant to cost-aware model routing.
Some ideas for what comes next, May 2026 — Nathan Lambert's roundup: Gemini Flash 3.5, open-vs-closed shift, economics of AI compute.

Better tools made Copilot code review worse — GitHub found migrating Copilot review to Unix-style exploration tools regressed quality until they reshaped agent workflow around PR evidence; cautionary tale for tool-choice in review agents.
Automating cross-repo documentation with GitHub Agentic Workflows — Aspire team auto-generates SME-reviewed doc PRs from merged product changes.
alibaba/open-code-review — Alibaba's internal review tool open-sourced: deterministic pipeline + LLM agent, line-level comments, fine-tuned ruleset (NPE, XSS, SQLi), OpenAI/Anthropic compatible.

SnapState — persistent state layer for AI agent workflows, HN launch.
diegosouzapw/OmniRoute — free MIT AI gateway aggregating 290+ providers/500+ models with quota-aware fallback and claimed 15-95% token compression via "RTK+Caveman"; check claims before trusting the compression numbers.
Why AI Infrastructure must evolve for Agent Experience — Modal CTO on infra built for agents vs humans, tied to Modal's $355M Series C.
CoreBunch/Instatic — self-hosted agentic CMS (Bun-based) outputting clean static pages, Webflow/Framer alternative.

Scriptc by Vercel — zero-runtime TypeScript-to-native compiler, no V8/Node in binary; community notes it only works for dependency-free greenfield code since real npm packages assume a JS runtime.
Claude driving Perslis MCP for legacy system control — MCP server claimed to bridge Claude into legacy code/systems control, unverified community post.
I run my project's marketing with Claude Code — memory layer implemented as a schema-validated markdown git repo, pattern worth stealing for agent state.
UseCaseVault — directory of AI workflows verified by people who actually ran them, built with Claude Code.
Show HN: GPT-4-powered web search for developers (Phind) — old-style dev search tool; community flags hallucination on empty results and unsustainable free-tier GPT-4 costs.