What makes an AI agent session valuable

Two sessions can be the same length and worth very different amounts. Here is what buyers look for.

#The short version

A valuable session is long, real and messy: a person with an actual goal, an agent using tools to reach it, things going wrong, and the agent working through them.

When Mercor advertised for coding sessions in October 2026, it asked for about 50 turns over two to three hours on a developer's own project. That is a good picture of the target.

#What raises value

Length with substance. Fifty turns of steering, feedback and correction show how a person and an agent actually work together. Ten turns rarely do.

Tools in the loop. Shell commands, file edits, test runs, MCP tools, browser actions. Labs train agents to act, and sessions without actions show only chat.

Errors and recovery. A failing test, a wrong assumption, a command that does not exist, then the fix. Recovery is the hardest thing for agents and the scarcest thing in training data.

A real goal. A feature someone needed, a bug that blocked a release. Invented tasks have a recognizable flatness, and buyers check for them.

Named products. For developer-tool companies, the value is in sessions that touch their product: setting up Supabase auth, deploying to Vercel, wiring PostHog events.

Uncommon ground. Less-covered languages, frameworks and domains: Elixir, embedded C, data pipelines, infrastructure work.

#What lowers value

• Secrets and personal data that would need heavy redaction.
• Code you do not own. Sessions on an employer's or client's codebase are usually not yours to license.
• Very short sessions - a single question, a one-line fix.
• Repetition. Ten sessions doing the same thing are worth little more than one.
• Generated sessions. Agents run against made-up busywork to produce data. Buyers screen for them.

#Check your own

The evaluator sorts your sessions into tiers by turns, active time and tool use, which is a rough proxy for the points above. Then look at your deep sessions yourself: which were real work, which were yours, and which you would be comfortable showing someone after redaction.