Let your AI agent switch models mid-task
Hey, vibe coders, an agent that gets stuck may need a different model, not another retry with the same one. This edition covers routing that adapts during a session, plus a reusable way to draft content from your own business experience.
In today’s edition
- OpenRouter adds model routing that adapts mid-session.
OpenRouter adapts model choices as agents work
OpenRouter's November 6 router update adjusts its model choices as an agent session grows longer or more complex. It also passes along signs that a model is stuck, helping the router decide when to switch.
- In OpenRouter's beta, the new router produced 4× fewer invalid tool calls and 50% fewer cut-off agent responses than the previous router.
- Latency was 3× lower than with the previous router.
- Auto Router covers general questions and agents; Auto Code targets coding agents. Optional API router metadata explains which models were weighed and why one was chosen.
What it means Try it if your agent stalls or makes invalid tool calls. Enable router metadata to see why it chooses each model.
Source: OpenRouter
Quick hits
- OpenAI and the Frontier Model Forum developed a framework with member firms to compare cyber jailbreak severity and prioritize stronger safeguards.
- OpenAI introduced BioBasicsBench with Ligo, using 100 real-experiment tasks to test agents' lab-data interpretation, error detection, and recognition of insufficient evidence.
GitHub repo of the day
Skills: give Claude repeatable document tasks
- What it does: Anthropic's repository provides reusable instructions and scripts for document work, including PDFs, spreadsheets, slides, and Word files. The maintainers include the document skills used in Claude as references you can adapt.
- How to start: In Claude Code, run /plugin marketplace add anthropics/skills, then /plugin install document-skills@anthropic-agent-skills. For a first task, ask Claude to use the PDF skill to extract form fields from a PDF you provide.
AI workflow of the day
How to draft posts from your business experience
Shared by Prajwal Tomar (@PrajwalTomar_) · View original post ↗
For creators whose AI drafts sound like everyone else's, this method separates research, story angles, writing, and editing. You supply the business experience and choose the hook; AI helps develop the draft using your past posts.
- Give the researcher role one job: find stories gaining attention in your niche over the last 48 hours.
- Have the chief role turn each story into angles connected to what you're actually running in your businesses.
- Give the writer your best past posts. Request 8 to 10 hooks, choose one, then have it draft the post.
- Ask the editor to check every fact and cut wording that sounds like generic AI writing.
- Keep rejected lines in a recurring writing-avoidance list for later drafts.