2025-12-12_pulse.md 10 KB
Pulse · December 12, 2025
What happens when your coding agent forgets everything you did yesterday? You start over. You re-explain context. You watch it make the same mistakes. It's the dirty secret of agentic coding—brilliant in the moment, amnesiac by morning. But here's the thing: the smartest people building these tools have been quietly solving this exact problem, and this week they finally started showing their work.
Anthropic dropped a piece on long-running agent harnesses that reads like the missing manual we've all been waiting for. The community responded with a flurry of parallel workflow tools, worktree managers, and at least three "awesome" lists that are actually awesome. Meanwhile, the practitioners who've been living in Claude Code for months are starting to share patterns that feel less like tips and more like battle-tested doctrine. The ecosystem is growing up fast—and if you're not paying attention, you're already behind.
The Signal
Effective harnesses for long-running agents
Anthropic Engineering · November 26, 2025
Here's the thing about agents: they're brilliant within a conversation, but ask them to remember what happened yesterday and you're back to square one. The Anthropic team has been wrestling with this exact problem—how do you build harnesses that let agents work across sessions without losing their minds?
Their answer draws on an unexpected source: how human engineers manage context when they step away from a problem. The patterns here (checkpointing, resumable state, explicit handoff protocols) feel like the missing manual for anyone building tools that need to survive a context window reset. It's the kind of deep thinking that usually stays inside research labs, now published for the rest of us to steal.
Pulse insights: This speaks directly to our /saveplan and /loadplan commands. The checkpoint pattern they describe is almost exactly what we're doing with session-cache.json, but they've thought through edge cases we haven't—particularly around what to do when the world changes between sessions. Worth a close read.
Compound Engineering: How Every Codes With Agents
Every · December 11, 2025
"What happens to software engineering when 100 percent of your code is written by agents?" Dan Shipper and Kieran Klaassen aren't asking hypothetically—they're running five production products at Every where nobody types code manually anymore. The piece introduces "compound engineering," where each feature makes the next one easier because the system learns from every bug, test, and breakthrough.
The real gem is their four-step loop: Plan → Work → Review → Compound. That last step—feeding lessons back into the system—is where most teams fall short. They've open-sourced a Claude Code plugin with the exact workflow, which is unusually generous.
Pulse insights: Their "plan first, code second" philosophy aligns perfectly with Claude Code's Plan Mode. The compound step is what we're missing—we should think about how /pulse itself could feed insights back into our agents automatically.
Official Updates
Advanced tool use on the Claude Developer Platform
Anthropic Engineering · November 24, 2025
Three new beta features that let Claude discover and execute tools dynamically. If you've ever wished Claude could figure out which tools exist without you listing them all upfront, this is the paper to read. The implications for skill discovery are significant—imagine Claude browsing your .claude/skills/ directory and choosing the right one without explicit triggers.
Pulse insights: Could change how we structure skill descriptions. If Claude can discover tools dynamically, our trigger keywords might matter less than clear capability descriptions.
Code execution with MCP: Building more efficient agents
Anthropic Engineering · November 4, 2025
Every tool call eats context tokens—the definition goes in, the result comes back, and suddenly you've burned through half your window on boilerplate. This piece explores how MCP code execution can compress those round-trips, essentially letting the agent "think in code" rather than "think in API calls." The efficiency gains are substantial for agents doing repetitive operations.
Pulse insights: Our fetch-expert agent would benefit from this pattern. Instead of multiple WebFetch calls, a single code execution could handle the whole parallel fetch loop.
anthropics/claude-code: Remove footer from code-review plugin
anthropics/claude-code · December 11, 2025
A small commit, but a telling one: the official code-review plugin no longer appends "Generated with Claude Code" footers. The implication? Anthropic is moving toward outputs that blend seamlessly rather than announcing their AI origins.
Pulse insights: We've been adding similar footers to our commands. Worth considering whether this is still the right call, or if we should follow Anthropic's lead toward invisible assistance.
GitHub Discoveries
The claude-code topic on GitHub has been busy. Here are the repos that caught our eye:
worktrunk
max-sixty · Git worktree management for parallel agents
You know how painful it is when multiple Claude sessions try to touch the same files? Worktrunk solves this with proper Git worktree isolation—each agent gets its own working directory, commits happen cleanly, and merges don't turn into nightmares. The CLI is minimal and focused.
Pulse insights: This is the missing piece for anyone running concurrent agents. We should experiment with integrating worktree patterns into our multi-agent workflows.
riffrag
naxoc · Local RAG + skills creator
A clever combination: build a local RAG index of your codebase, then use it to generate Claude Code skills automatically. The idea is that your code patterns become teachable skills without manual documentation. Still early, but the approach is sound.
Pulse insights: The auto-skills-from-code concept could supercharge /agent-genesis. Instead of manually describing expertise, let the code speak for itself.
spec-oxide
marconae · Spec-driven development with MCP
Written in Rust (naturally), this tool enforces spec-first development where you write requirements, the agent implements, and specs become the source of truth. The MCP integration means Claude Code can read and validate against specs natively.
Pulse insights: Spec-driven development + MCP is an interesting pattern for complex features. Could inform how we structure /plan outputs.
Claude-Multi-Agent-Research-System-Skill
zubayer0077 · Orchestrated research agents
A skill for running multiple agents in parallel on research tasks, with built-in audit trails. The orchestration layer handles task distribution and result synthesis—essentially a conductor pattern for agent swarms.
Pulse insights: The parallel execution model here is more sophisticated than our current approach. Worth studying for potential fetch-expert improvements.
Community Radar
SkillsMP — The skills keep coming
The Claude Skills Marketplace now indexes hundreds of community skills, sortable by stars and categories. Quality varies wildly, but it's becoming the de facto discovery layer for anyone not wanting to reinvent wheels.
Simon Willison — Consistently essential
Simon's been on a tear with MCP coverage. If you're not subscribed, you're missing the clearest explanations of what these tools actually do (and don't do). His "things I learned" format is worth stealing.
Shrivu's Substack — Practitioner dispatches
Real workflows from someone actually shipping with Claude Code daily. Less theory, more "here's what broke and how I fixed it." The kind of content that's hard to find elsewhere.
Quick Hits
- Awesome Claude AI: Fresh MCP server listings and SDK integrations—the directory's getting genuinely useful
- vibeproxy: macOS menu bar app for using Claude subscriptions without API keys. Yes, it works.
- claude-code-mux: Multi-model routing in single sessions—OpenAI and Claude playing nice together
- mcp-memory: Persistent memory layer for MCP—agents that actually remember things between sessions
- claude-code-metrics: Official usage analytics for tracking token spend and session patterns
- cursorrules.io: Growing collection of
.cursorrulesfiles—steal patterns for your.claude/rules/
The Hit List
- Read the harnesses article — Anthropic's checkpoint patterns are worth stealing for
/saveplan - Clone compound engineering — Every's plugin is free, battle-tested, and solves problems you haven't hit yet
- Set up worktrunk — Your parallel agent workflow is a mess without proper worktree isolation
- Kill your footers — Anthropic removed "Generated with Claude Code" from their plugins. Take the hint.
- Audit your skills — Half of them probably duplicate built-in functionality now
Sifted by Pulse · 12th December 2025 · 16 tabs so you don't have to