Most important take away

The $285 billion SaaS stock selloff was triggered by AI agents like Cowork that are still in research preview with fundamental limitations (e.g., stops working when you close your laptop). When evaluating any AI agent tool, ask three questions: does it have persistent memory, does it produce editable artifacts, and does its context compound over time? Even Cowork only scores about 1.25 out of 3.

Summary

Actionable Insights

1. Three evaluation questions for any AI agent: (1) Does it have persistent memory? (2) Does it produce inspectable/editable artifacts? (3) Does context compound over time? Use these before buying or building.

2. Build vs. buy consideration — You can build your own agent infrastructure for ~$0.10/month using open-source tools (like Open Brain) instead of $20-200/month SaaS agent tools, if you’re willing to invest the setup time.

3. Agent tool ratings:

   • Cowork (Anthropic/Microsoft): Best artifact production, weak on memory persistence and context compounding. ~1.25/3.
   • Lindy: Easy onboarding for executives but opaque artifacts, credit burn issues, 2.4/5 on TrustPilot. Mixed.
   • Sauna (pivoted from Wordware): Strongest memory architecture conceptually, but very early/demo-heavy. Wait and see.
   • Google Opal: Free, good community energy, remixable workflows, but memory is spreadsheet-simple and Google may abandon it.
   • Obvious: Most ambitious (full workspace with SQL, docs, presentations, kanban), but too new to evaluate.

4. Three-layer architecture for DIY agents: Knowledge store (memory/database), Recipes (pre-wired workflows), and Scheduling loop (run schedules with learning over time).

Career Advice

• The future of agent work requires understanding verifiable domains — code was first because you can easily tell if it works. Non-code domains are harder to verify, which is why agent quality varies.
• Knowledge workers won’t become programmers but need to write clear specs — clarity of intent is the key skill.

Chapter Summaries

1. Why Outcome Agents Exist — Cowork launched autonomous computer-using agents, Microsoft copied it, and $285B left SaaS stocks. But Cowork is still in research preview with basic limitations.

2. Three Evaluation Questions — Framework for judging agents: persistent memory, editable artifacts, compounding context. Even Cowork only partially meets these criteria.

3. Agent Reviews — Detailed assessment of Lindy (executive-focused, mixed results), Sauna/Wordware (promising memory architecture, too early), Google Opal (free, community-driven, shallow memory), and Obvious (ambitious workspace, too new).

4. Three Principles & DIY Architecture — Memory must be architectural, outcomes need editable surfaces, context must compound. Three-layer architecture: knowledge store, recipes, scheduling loop.