CEREBRO

Coding agents & orchestration

Orca runs Codex, Claude Code, OpenCode, and Pi side-by-side in separate git worktrees under one ADE, useful if you're juggling multiple agent runs on the same repo.
book-to-skill converts any PDF/doc collection into a Claude Code skill, claiming 24-51x fewer tokens than dumping the source into context per query.
agent-skills packages senior-engineer workflows (define/plan/build/verify/review/ship) as reusable Claude Code skills for consistency across sessions.
archify is a skill that turns a codebase or system description into self-contained interactive HTML architecture/sequence/flow diagrams.
Munder Difflin lets you spin up clones of yourself running Claude Code and Codex to offload work.

Claude Code v2.1.233 adds GitLab MR support to --worktree and a forward_user_identity gateway setting for per-user spend attribution behind a proxy.
Maximizing the value of your Claude Code sessions covers /clear between tasks and /context for token hygiene, though HN pushback notes unexplained cache blowups (400K→2M tokens) that the post frames as a user problem rather than a tool bug.
A privacy-first memory layer for Claude multi-agent fleets is a community build worth a skim if you're evaluating memory-layer alternatives to your current setup.

Opus 5: Fable-level performance at half Fable's price beats Fable on most official benchmarks while Anthropic still undersells it as "comes close," reflecting eval difficulty more than capability.
Why does Opus 5 feel worse to work with? argues Opus 5 is more capable than 4.7/4.8 on paper but feels worse day-to-day: slower, more obtuse, over-commented, and prone to using cached data instead of computing live.
Claude Fable 5 and new AI safety fables tracks Anthropic's shift to classifier-based handling of "safety" flags on AI research queries.
Gemini 3.7 Flash is pitched as Google's strongest workhorse model yet for coding and agentic tasks.
GLM-5.2 marks a capability threshold for open-weight agents, paired with a new paper on open RL recipes for terminal agents.
Qwen 3.8 27B reportedly edges out Opus 4.7 Max on DeepSWE (42.2 vs 40) while running on consumer hardware.
GPT-5.6 Sol previews as 14x faster, aimed at real-time agent workloads.
Codex from 0 to 10M Users traces Codex's >10x MAU growth since January, framed around agentic interfaces for the much larger population that can't write code but can direct it.

Grok 4.6 now in GitHub Copilot targets agentic coding and multi-step workflows, per GitHub's internal testing.
GitHub Copilot weekly releases, Aug 10 rounds up new models, portable plugins, and smoother agent workflows across editors and CLI.
Bringing your software delivery workflow into GitHub with agent apps shows four GitHub agent apps covering scope/secure/rollout/ship without leaving GitHub.

OfficeCLI gives agents a single-binary, no-dependency way to read/edit/automate Word/Excel/PowerPoint via an HTML rendering engine.
SnapState offers persistent state specifically for AI agent workflows.
OmniRoute is a free MIT gateway aggregating 290+ providers/500+ models behind one endpoint with quota-aware fallback and claimed 15-95% token compression.
TencentDB Agent Memory is a team-level memory hub turning chats/docs/code into shared, governed memory assets across agents.
ego-lite is a browser built for AI agents to share your logged-in session state without disrupting your own tabs, pitched against browser-use/agent-browser style automation frameworks.

Mole is a terminal deep-research agent with enforced per-call budgets, quote verification, and contradiction-checking across sources; HN flags the codebase as bloated for what it does and budget/caching interaction as unclear.