There's a wide gap between "works" and "works well." The difference shows up in your token bill.
Your agent team ships. The output is correct. The pipeline doesn't break.
It costs too much. It runs too slow. It burns tokens on work that a cheaper model could handle. It runs steps sequentially that could run in parallel. It skips caching that would cut redundant calls by half.
The inefficiencies compound. A 2x token waste on a $2 build is noise. A 2x token waste on a production pipeline running daily is a budget line item you can't defend.
The Tuning Pipeline. Audit the pipeline. Set targets. Rewrite the config. Measure the difference.
The Coach reads every agent call in the pipeline. Which model handles which task. Where tokens go. Where time goes. Where waste accumulates.
Speed, cost, quality, reliability. Pick the two that matter most. You can't optimise for everything at once. The Coach helps you pick.
Model swaps. Prompt compression. Caching layers. Parallel execution paths. The factory applies the changes to your existing pipeline, not a greenfield rewrite.
The factory re-runs the pipeline with the same inputs. You get the numbers side by side. Token count before, token count after. Duration before, duration after.
Optimisation data comes from aggregate factory operations. Patterns across 23 builds, not a single flagship.
The factory shipped with 112 unit tests passing. The skipped phases were rate-limited, not broken. The pipeline knew when to stop retrying and move on.
"Tell us what you need optimised."
Depends on the pipeline. Some teams cut token costs by 40% with a model swap. Others cut duration by half with parallel execution. The audit gives you the specific numbers before you commit.
Token count, duration, cost per run, failure rate, and the ratio between useful output and overhead. The audit produces a baseline. The rewrite produces a comparison.
The Coach signs an NDA. The audit traces the pipeline without exposing it. The before/after numbers are yours to keep regardless.
Yes. The pipeline doesn't care who wrote it. The audit reads it the same way either way.