One signal: one agent beat four

One agent, one job

The best agent build published this week was one agent doing one recurring job. New research explains why that beats a swarm.

Anthropic showed how its own field marketer built a weekly digest for every sales rep. Claude Code reads BigQuery through MCP across HubSpot, Clay and Salesforce, pulls each rep's territory from Slack, and sends each one a Monday message with three priority actions. Registrations for one executive dinner doubled in a week. Hours of Sunday work disappeared.

The unglamorous parts are why it still runs: nine content rules written from real feedback, column-header validation so a schema change cannot silently break it, every send archived.

Now the research. Exponential View's read of Anthropic's multi-agent study found that on problems where the answer depends on information only one agent holds, most four-agent teams got it right 17 to 36 percent of the time. One agent given the complete evidence got it right nearly every time. The failure was sameness. Eighteen of thirty agents named their git branches identically.

What this means for you: take one job you redo every week. Give one agent all of the context. Put your effort into the rules and the checks, not into adding agents.  → the build  → the research

Thomson Reuters built its own frontier model for $40 million. Trained from an open-source base on decades of Westlaw and Reuters archives, using under 10% of that content so far. The company says early evaluations put it on par with the latest frontier models, and independent scrutiny of those benchmarks has already started. What this means for you: the model was the cheap part. The archive was the asset.  → the announcement

Claude's memory now works across chat and Cowork. It writes what it learns as you talk rather than summarizing at the end, and every stored topic is listed under Settings, where you can edit or delete any of them. Health and beliefs stay out by default. What this means for you: go read what it already knows about you, and delete what is wrong.  → what changed

Unlocked Today is free. Swept by Nova, our AI news agent; edited down to what matters. Two emails a week, Sunday and Wednesday, and nothing else.

Past editions: go-unlocked.com/newsletter

Reply anytime. I read them.
Kendall

You're getting this because you subscribed at go-unlocked.com.  Unsubscribe