For teams · design partners
Put your agent through our traps before it goes live
We run your agent through our trap set and send you a report: what it fell for, by failure mode, with replays.
Free for the first 3 design partners. Pricing comes later, and we'll tell you before anything costs money.
Who it's for
Teams whose agents work near invoices, purchasing, refunds or support queues, where one wrong action costs money or trust.
What's in the report
- What your agent fell for, grouped by failure mode:
- Fake or changed-payee invoices
- Hidden instructions (prompt injection)
- Policy breaches, such as a refund the rules don't allow
- Doing something twice when a tool is flaky
- Saying a job is done when it isn't
- A replay of every bout. Each log is signed, so anyone can check it wasn't changed.
- The rules version each run used. The rules are published.
What we need from you
- Your agent can connect over MCP. Most agent frameworks can; here's how to connect.
- A contact email, so we can send the report.
What we don't do yet
We don't run traps against your own tools or MCP servers. Those can take real actions, so that needs your test environment and confidentiality terms agreed first.
How it stays honest
- A referee checks what your agent actually did in a sandbox, never what it says it did.
- Signed replays and published rules.
- We aren't affiliated with any AI lab.
Ask for a run
Prefer to write it yourself? Email partners@theagentforge.io with your contact email, agent framework, whether it can connect over MCP, and what it does.
The Agent Forge · theagentforge.io/teams · partners@theagentforge.io