← All capabilities
Optimisation
Phase: ImproveLower the cost and raise the reliability of what’s already in production.
Optimisation improves what already works: less cost per case, more reliability, less human intervention. It leans on observability data and evals to change with judgement, not hunches. It’s what turns a profitable pilot into an efficient operation at scale.
What it includes
- Cut cost per case: a cheaper model where it suffices, fewer steps, caching.
- Raise reliability where it hurts most, guided by production data.
- Reduce human intervention by expanding what the agent resolves alone with confidence.
- Close the loop: every production failure enters the eval and improves the next version.
When you need it
- The process works but cost per case is too high to scale.
- An agent resolves, but escalates to a human more than it should.
- You want to take the human out of the loop and need to gain reliability first.
How it’s measured
- Cost per case over time (trending down).
- Resolution rate with no human intervention (trending up).
- Reliability per process measured against the eval.
Common mistakes
- Optimising on a hunch with no observability data or evals to back the change.
- Cutting cost at the expense of reliability without measuring the damage.
- Optimising a process that should be switched off because it returns nothing.
FAQ
- Is optimisation the same as maintenance?
- No. Maintenance holds reliability against external change; optimisation actively improves the cost and performance of what already works. One defends, the other pushes upward.
- When do you optimise an agent?
- Once it works and there’s enough production data to know what to improve. Optimising before you have reliability and measurement is polishing something you may have to rebuild.
Related capabilities
Missing any of these?
Start with an assessment: which capabilities you have, which you’re missing, and where it pays off most to start.
Talk to us →