WORTH TESTING · 23 OCTOBER 2026
Anthropic's Jess Yan on building long-running agents (a manipulated re-upload: watch Peter Yang's original instead)
Watched in full (40 minutes). This is a re-upload of Peter Yang's podcast interview with Jess Yan, introduced as Anthropic's product lead for Claude Managed Agents. The audio and demos look genuine, but the faces appear to be AI-generated avatars, and the post's quote about '90% of engineers using Opus 5.5 and self-improving loops' is never said in the video. Verdict: great ideas on harnesses, self-correcting loops, evals and starting small; intermediate to advanced, and you need coding for tools and MCP. Watch the original on Peter Yang's YouTube channel, linked below. It has ten things to test.
Try this
- Think of an agent as model plus harness: the harness handles tools, memory and permissions. 02:00
- Break the goals in your system prompt into numbered, separate steps for more predictable results. 09:00
- Use MCP as the secure, signed-in way for an agent to reach your internal data. 10:17
- Design loops that read their own errors and retry, rather than fragile chains of steps. 11:46
- Use a separate AI as a judge to grade the agent's output. 14:18
- Run checks at the halfway points of long tasks, not only at the end. 15:18
- Let agents produce a finished web page or interactive component instead of rigid data formats. 17:00
- Connect an agent to code reviews so it checks every change. 19:45
- Test by hand ('vibe-test') before building formal evaluation pipelines. 31:45
- Start with one specialised agent, not a 20-agent system. 34:04
Timestamps point to the moment in the source video.
Cost
Free to watch.
Source
Hype or Fact?
Hype A 'manipulated re-upload' of Jess Yan on long-running agents
Related AI tools
How we knowSteps to try, not a claim that we independently tested every result. Follow the linked sources and check current prices. Our methods & sources ↗