This week I wired two OpenClaw agents together across two VPS instances over Tailscale and spent an evening as the dumbest node in my own network — copying messages between them like a human packet router. That one session broke open a whole category of problems: agent-to-agent communication, identity vs. tokens, async callbacks, and why Discord is the wrong message bus. Twelve articles came out of it. Here's the first weekly roundup.

Connecting Two OpenClaw Agents Over Tailscale: How ACP Pairing Works
How to wire agent-to-agent communication in OpenClaw using ACP, and why the pairing-required loop has nothing to do with your token.
openclaw · integration

Ditch Discord as Your Agent Message Bus
Replace Discord with direct Tailscale HTTP for agent-to-agent communication and inject messages into a remote OpenClaw agent.
openclaw · pattern

Your Agent Inbox Needs a Callback
HTTP endpoints that trigger LLM turns can't block waiting for a response. Accept immediately, process async, give the agent a route to reply.
openclaw · integration

Route Expensive Model Calls to Subagents with sessions_spawn
Use sessions_spawn's model parameter to run heavy tasks on Opus while keeping your orchestrator on a fast, cheap model.
openclaw · pattern

When to Farm and When to Just Do the Work
Most orchestrator agents default to spawning subagents for everything. The overhead only pays off on tasks that won't fit in one session.
skills · pattern

What a Skill Audit Actually Finds
Auditing 20 agent skills surfaced undertriggering descriptions, missing persistence, stale docs, and duplicated boilerplate. Here's the pattern.
skills · pattern

Turn Your Agent's Standing Procedures Into Skills
Extracting workflows buried in config docs into structured skills makes them triggerable, testable, and independently improvable.
skills · pattern · 1-2 hours

Edit Tool Fails the Second Time the Same Field Gets Updated — Use sed Instead
The Edit tool fails when old_string no longer exists. For dates, timestamps, and counters that update repeatedly, sed is the right tool.
skills · fix · 5 min

Plus 4 more on the site: agents batching signals silently, nodes vs. messaging, system cron failures, and prompt-as-bug.

📊 This week: 12 new fixes · 17 total in the library · Most active: multi-agent communication over Tailscale · Frameworks: OpenClaw, Skills

— Adam Kinney, building agents at aiallthethings.com