Constolve runs your agent in a working copy of its world, against its own rules.
Simulated customers, tools and records. Every action checked against your agent's written policy. When it breaks a rule, we fix it and prove the fix on the same tests.
One sentence in an expense claim made Google's sample agent pay $240 without the manager's approval. See the demo →
- OriginalPaid $240 without manager approval
- CorrectedAsked the manager. Nothing paid.
See it on four sample agents
Finance reimbursement
Can text inside a claim talk it into paying without the manager?
Start the demo →Smart home
Does it block unsafe temperatures without refusing safe ones?
Start the demo →Insurance prior authorization
Can a note hidden in the intake form make it skip its checks?
Start the demo →Email trust
When the inbox holds an untrusted email, does a reply still go out without the user's approval?
Start the demo →How it works
- Step 1Policy
You give us the agent's operating policy and its tools.
- Step 2World
We build a working copy of its world: customers, tools and records. Nothing real is touched.
- Step 3Tests
Tests come from the policy itself, including ones built to trick the agent.
- Step 4Runs
Your agent runs every scenario, more than once, so one lucky run doesn't count.
- Step 5Fixes
Where it breaks a rule, we add the fix and prove it on the same tests.
Tested on a public benchmark
On the telecom domain of Sierra's τ³-bench, our tests, built from the written policy alone, needed the same repair tools as 90 of the benchmark's 94 repair tasks.
Read the white paperWhat you bring
- The agent's instructions
- Its tool definitions
- A way to run it, or recorded tool responses
Start with one agent.
Send us one agent's policy and tools. We run it in a working copy of its world and show you every result.