From the source
Together AI says small fixes to its docs, skills and API errors raised coding-agent task completion on its platform from 75% to 97%.
- The company built an agent evaluation suite after noticing more developers were pointing coding agents at Together AI, and it had no way to tell whether those agents could work with the platform efficiently.
- Each task is run three ways: a baseline with only the prompt, a configuration with direct documentation access through Together's MCP server, and one given the Together skill package of agent-written instructions and examples.
- Every task runs against three models with two attempts each, giving six attempts per task per configuration, and agents run in a Together sandbox with
GitHub Actions handling setup and cleanup while a proxy records API calls.
- An LLM judge scores each run as pass, incomplete or fail, and Together says failed and incomplete verdicts only show where to start looking, not whether the fault was the agent, the API or missing evidence.
- Together also reviews passing runs, since agents that work around confusing responses reveal what to fix for token efficiency.






