AgentFence after the WebMCP Challenge: trace-backed security receipts
The hackathon submission is frozen, but we're continuing to explore where AgentFence can go beyond the original MVP.
The next step is observability.
We added OpenTelemetry instrumentation around the AgentFence authorization lifecycle so a consequential agent action can now be correlated across:
WebMCP tool execution → provenance → security evidence → policy evaluation → human approval → mutation → verification
The biggest addition is the trace-backed security receipt.
Instead of a receipt existing only as a static record of an approval, it now carries the OpenTelemetry Trace ID for the complete remediation lifecycle.
For example:
TAINTED provenance → CRITICAL action → human approval → apply_fix → verification → PASS
can all be correlated through a single execution trace.
This gives us a path toward reconstructing not just what an agent did, but:
- what capability it invoked
- what provenance the action carried
- what security evidence was available
- what policy decision was made
- whether human authorization was required
- what mutation occurred
- whether verification passed
We're deliberately keeping raw repository content and prompt text out of telemetry. The goal is an auditable security trail, not another copy of application data.
The original WebMCP Challenge submission remains unchanged and frozen. This is the direction we're exploring next for AgentFence.
WebMCP gives agents capabilities. AgentFence makes those capabilities governable.
And now we're exploring how to make that governance observable and auditable.
Log in or sign up for Devpost to join the conversation.