Put checks in the right place.
Investigate whether verification belongs before a state change. Test the proposed reorder against the relevant cases.
Agent optimization / with Cave Agent
Find the waste.
Fix the workflow.
Move from costly traces to reproducible cases, evaluated changes, and scoped draft proposals. Give your agents a feedback loop—not another dashboard to ignore.
Caveman Platform · private development · human-reviewed changes
Finding the expensive run is the beginning.
A chart cannot
fix your agent.
01 / From evidence to a reviewable change
A smaller fix.
A stronger case.
Cave Agent is the proposal step in the optimization loop. Follow the evidence, test the intervention, and review a bounded change with its source and scope attached.
Draft proposal / no auto-merge
A fix you can review.
Cave Agent turns an approved, bounded change plan into a draft proposal. File limits, allowed commands, and source provenance travel with it.
Your scope. Your review. Your merge.
DraftVerify before changing state
agent/workflow.ts +2 −2
async function updateRecord(input) {
− await write(input);
− await verify(input);
+ await verify(input);
+ await write(input);
} Bounded file scope Evidence linked○ Human review pending
Conceptual diff, not a generated patch · proposals do not merge themselves or create verified savings
02 / Diagnose the path, not just the model
Sometimes the problem
is the workflow.
A retry loop, a misplaced check, or a missing guard can matter more than the model price. Inspect related paths and the evidence behind each opportunity.
Repeated verification
01Read
02Verify
03Read
04Verify
05Answer
The same artifact gets read and verified twice. Inspect the traces before deciding whether either check is redundant.
Illustrative pattern analysis · themes require suitable captured evidence · lexical clusters are not labeled semantic
Guard the exception.
Add an applicability guard or a guarded path for the work that needs it. Keep the intervention bounded by the diagnosed case.
Keep the reasoning inspectable.
Connect an opportunity to its traces, task family, diagnosis, and proposed intervention. Let reviewers follow the chain.
03 / A proposal with boundaries
Fixes need scope.
Not blind trust.
Approved inputs become bounded proposals. Review the file and command limits alongside the evidence before any change enters your repository.
Read records
Write recordsNot granted
Delete recordsNot granted
One case. A bounded change.
A change plan declares allowed files, change limits, commands, and network policy. The stricter harness limits still apply.
Quality
Cost
PolicyPending
Tests travel with the idea.
Build an eval pack around the diagnosed behavior. Recorded replay informs review; it does not stand in for a live canary.
Run receipt
Model calls2
Tool calls3
Stop reasonCompleted
Illustrative receiptYou keep the merge button.
Cave Agent produces proposal-only drafts. A draft cannot auto-merge or mint verified savings.
Illustrative controls. Actual proposal eligibility depends on supported interventions, approved inputs, repository binding, and evidence readiness.
04 / One optimization workspace
Find. Test. Review.
Keep improving.
Opportunities, running optimizations, and evaluation evidence share a workspace. These are real product screens from a seeded demo.
● ● ●Demo workspace
Real product UI · seeded demo workspace · catalog list-price subtotals
05 / Clear expectations
Is Cave Agent a general-purpose bug-fixing bot?
Its role here is narrower: convert approved optimization inputs into scoped draft proposals. Supported workflow interventions include guarded paths, applicability guards, and verification relocation. It is not a promise to fix every bug in a repository.
Will it merge changes automatically?
No. This proposal path is draft-only. Repository scope, evaluation evidence, and human review remain part of the decision.
When do improvements count as savings?
An opportunity, an estimate, a replay result, and a draft PR are not verified savings. Savings require the appropriate measured, provider-causal evidence after an intervention.
Make every token count.
Give your agents
a better next run.
Connect observation, evaluation, and review. Make improvements you can explain—and changes your team can own.