Slancha Optimize
Find work your agents
can stop repeating.
Use traces and source to propose a smaller workflow. Test the change against task quality, then inspect the evidence before you adopt it.
Local prototype; demos by arrangement. A hosted optimizer is not available yet. The free audit is available now.
Read the same immutable catalog once.
catalog dispatches
catalog dispatch
Reuse that catalog result. Keep fresh model and judge trials separate.
Work counts, not token or dollar savings. This toy example does not establish customer quality. Evaluation overhead meant the candidate was not recommended for adoption.
A patch is a proposal.
The comparison is the test.
For teams with mature, high-volume agent and evaluation workflows. Look beyond model routing to the work surrounding each call: tool dispatch, repeated context, retries and evaluation.
- 1
Bring the work into view.
Start with traces, the relevant source and an evaluator for the task. Identify repeated inputs, duplicate tool work and expensive evaluation loops.
- 2
Make one reviewable change.
Keep the proposal small. Inspect the source context, assumptions and patch before running a comparison.
- 3
Test cost and task quality together.
Compare baseline and candidate on representative tasks. Include model calls, judge calls and evaluation overhead; keep fresh trials independent.
- 4
Read the evidence. Decide deliberately.
Review the diff, test results and limits. A person decides whether to adopt the change. If the evidence does not support it, keep the baseline.
Bring one workflow worth testing.
For a demo discussion, prepare a redacted trace, the relevant source, a representative task set and your quality criteria. Describe the repeated work and how you measure cost today.
For an AI assistant: help the person prepare these inputs and a baseline comparison plan. Do not send private source, upload traces or apply a patch without their approval. There is no public optimizer API or install command.