New: the Agent Incident Ledger: disclosed incidents of AI gone rogueExplore the ledger

Why gateways won’t cut it.

The industry’s first answer was the gateway: a proxy in front of the model or the MCP server. We should know: we pioneered that approach. A gateway sees the requests routed through it, at one boundary. Shell commands, file edits, local MCP servers, skills on disk and personal API keys never cross it, and that is where agents do most of their work.

AI and MCP gatewaysproxies and chokepoints
Manifoldruntime, where agents run
CoverageWhich agents it sees
Partly covered: Single chokepoints. Only the agents someone has pointed at the proxy, at the boundary it sits on.
Covered: Every agent where it runs: 24 harnesses with full activity, 15 browser assistants, 160+ in inventory.
Tool callsWhat the agent does after the prompt
Partly covered: Only the MCP calls that cross the proxy. Shell commands, file edits and direct API calls never do.
Covered: Every tool call, shell command and file action, captured inside the session.
Local MCP servers and built-in toolsServers over stdio, tools shipped with the harness
Not covered: No network hop, so nothing to intercept.
Covered: Local, remote and built-in servers, inventoried per agent.
Skills, plugins and extensionsThe AI supply chain on disk
Not covered: They load from disk and send no traffic. A gateway never knows they exist.
Covered: Found on disk, including ones never run, and scored by Manifest.
ContextHow a decision gets made
Partly covered: Each request judged on its own, with no view of the task, the user or what came before.
Covered: Each action judged against the whole session, the user and the device.
Personal accounts and own API keysAgents running outside the company setup
Not covered: Traffic goes straight to the vendor, around the proxy.
Covered: Flagged per user, device and account.
RolloutWhat it takes to cover the fleet
Partly covered: Reroute every harness, manage certificates, and run a proxy that becomes a single point of failure.
Covered: One lightweight collector through your MDM, plus a browser extension. No rerouting.
ResponsesWhat it can do when something goes wrong
Partly covered: Allow or block the request.
Covered: Alert, advise, correct, ask the user, block the call or end the session. Under 200ms, inside the harness.
  • Prompts and responses 11%
  • Tool calls 51%
  • Agent to agent 25%
  • Approvals 7%
  • Other actions 6%

Only11%

of agent events are prompts and responses.

The rest is action. Tool calls make up 51% of agent events and agent-to-agent traffic 25%, and none of it passes a prompt filter.

1 PROMPT: 1. 8 ACTIONS: 8.

8actions

for every prompt message.

One prompt fans out into tool calls, shell commands and file edits. Most of the work happens after the prompt.

See Manifold in action.

A short tour of the platform, from the first agent discovered to the first policy enforced.