AI NewsProductivityAnnouncement
Together AI lifted coding agent task completion from 75 to 97 percent
Together AI measured how many of its own tasks a coding agent could finish. Small fixes to docs, skills and API errors lifted the rate from 75 to 97 percent.

Image: Together AI
Why it mattersA docs change can move a coding agent more than a new model. Together AI's own eval shows what to fix and in what order.
A team building a tool that a coding agent uses writes docs for a human reader. The human then does not read them. The agent reads them, and the agent fails in places a person would not. Together AI measured where its own platform failed a coding agent. It then fixed the places and measured again. The pass rate went from 75 percent to 97 percent.
Together AI published the method and the numbers in a 16 November post. The authors are the Together AI team. The post names three models the eval ran against: Claude Sonnet 5.5, Claude Opus 5.5 and GPT-6.1 Sol. It does not say which brands of coding agent were tested beyond the model.
What the eval does
The eval runs 12 plus tasks through the platform. Each task runs on three configurations: a baseline, docs plus an MCP server, and a unified skills pack. Each configuration runs six times on each model. A sandbox tracks every API call the agent makes. An LLM judge reads the trace and marks each run as Pass, Incomplete or Fail. A small dashboard puts the results side by side.
The number Together AI publishes
The 75 percent is the baseline on 15 June. The 97 percent is the number on 8 October, with the unified skills pack. These are Together AI's own figures, measured on its own tasks. The post does not state the exact task list or the per-model split.
The fixes that moved the number
The post names three. A GPU cluster task was failing because the agent could not find the pricing page or the live regions API. Two sentences in the docs fixed it. The unified skill pack removed context bloat, corrected stale product information and fixed bugs in the example code. The API error message was fixed where it pointed to a path that no longer existed.
None of the three fixes is a model change. None is a prompt change. Each is a docs or an error-message change, found by watching one agent fail and then reading the trace.
For a team shipping an API
The method is reusable. Pick the 10 or 20 jobs an agent should finish on the platform. Run each on the models a user is likely to pick. Pass, Incomplete, Fail. Read the traces on Fail and Incomplete. Fix the thing the trace points at. The thing is almost always a docs sentence, a skill file or an error message, not a model.
Together AI says its docs and its skill pack carry the fixes that moved the number. A user on the platform gets the new behaviour without doing anything.
Source
Together AI: How we make Together AI easy for coding agents to use.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.

