CEREBRO

Claude Code Runtime

Sub-agents can now spawn sub-agents up to 5 levels deep — v2.1.172 also adds Bedrock region resolution from ~/.aws config and a plugin search bar.
v2.1.183 blocks destructive git ops in auto mode — git reset --hard, git clean -fd, git commit --amend (not agent-authored), and terraform destroy are now gated unless you explicitly asked to discard work.
Claude Code is the only agent shipping tool search on by default — the thread argues this architectural default is what makes MCP viable in practice vs. every other agent runtime.
Claude's WebSearch returns title+URL only; WebFetch routes 100 KB through Haiku 3.5 before the main model — citations capped at 125 chars, so there's a compression/summarization middle layer most prompts don't account for.

"It got dumber" is context rot, not model regression — follow-up post with 5 concrete fixes for degradation caused by accumulated tool output and contradictory turns.
Compaction is lossy — that's the actual root of "Claude forgets everything" — accepting this changes how you structure long sessions (checkpoint summaries, not prompt tweaks).
The single most costly token mistake — thread surfaces that re-sending large unchanged files each turn accounts for the bulk of wasted spend.
A model listed 78% cheaper cost 22% more to actually run — input/output ratio skew means per-token sticker price is unreliable; measure by workload output unit.
chopratejas/headroom — library/proxy/MCP server that compresses tool outputs, logs, files, and RAG chunks 60–95% before they hit the LLM, with no answer quality loss claimed.
DeusData/codebase-memory-mcp — tree-sitter AST-backed MCP server, single static binary, 158 languages, sub-ms structural queries, claims 99% token reduction vs. file-dump approaches; indexes Linux kernel in 3 min.

GitHub's plan for Agents — Kyle Daigle interview — Latent Space pod covering GitHub's agentic roadmap and Copilot positioning alongside the MS Build/Satya Nadella crossover.
Copilot improves context handling and model routing — explains how Copilot now trims context and routes to cheaper models for low-complexity turns to stretch per-seat credits.
Copilot CLI now more selective about agent delegation — orchestration heuristics tuned so fewer tasks hand off to slower sub-agents, no new config knobs.
Copilot CLI + LSP servers for real code intelligence — replaces grep/decompile context-building with proper language server symbols; setup guide included.
Copilot code review now reads AGENTS.md — repo-level AGENTS.md files give Copilot review its own instruction layer, plus a Request button for draft PRs.

Zero-Touch OAuth for MCP now stable — Enterprise-Managed Authorization extension lets orgs centrally manage MCP server auth; Anthropic, Microsoft, and Okta are adopting it so users get single-login access across all connected servers.
API to MCP — ProductHunt launch: wraps any REST API as an MCP server for agents without writing glue code.

addyosmani/agent-skills — production-grade skill definitions covering the full define→plan→build→verify→review→ship lifecycle, designed so agents follow senior-engineer workflows consistently.
obra/superpowers — composable skill + methodology framework for coding agents; hiring a community engineer, signals this is moving toward an ecosystem.
withastro/flue — TypeScript agent harness from the Astro team; imports SKILL.md files as typed modules (with { type: 'skill' }), interesting ergonomic pattern for skill-as-import.
BuilderIO/agent-native — framework where UI and agent share one database and state; actions work both by click and by prompt, real-time multiplayer between humans and agents.
SnapState — persistent state layer for agent workflows so state survives between runs without custom storage wiring.

kenn-io/agentsview — local-first single binary that aggregates session logs, token spend, and analytics across Claude Code, Codex, and 20+ other agents; no account required.
calesthio/OpenMontage — 12 pipelines, 52 tools, 500+ agent skills for agentic video production; worth watching as a large-scale real-world skills deployment example even if video production isn't the use case.