Transcript note: one short passage (on the GitHub-issues MCP) came out garbled; the surrounding context is intact.
The hosts walk through a personal-assistant agent (an OpenClaw-style “claw”) built around meetings. It tags calendar events by type, writes per-type morning briefings (with past action items and enrichment on external guests), records every meeting to Notion, writes templated summaries, files action items, drafts follow-up emails in the user’s style, and learns nightly from the user’s corrections. They compare it with n8n/Zapier (deterministic and cheap, but unnatural to build) and Anthropic’s dynamic workflows (converse first, then generate deterministic code). Team pain points: agents acting under the user’s identity flood shared tools with noise; skills shared by forking diverge; skills need versioning like libraries. A self-healing loop: users’ complaints become GitHub issues, an agent SDK proposes fixes, small ones auto-merge, and the user gets told to retry.
Show: langtalks
Concepts: personal-ai-assistants · skill-engineering · memory-consolidation · agent-workspaces · second-brain