Product
One platform for inventory, runtime control, approvals, and replay
AgentGuard is intentionally opinionated: it sits in the active path between agent context, decisions, tools, and external actions.
Agent inventory
Map tenants, agents, models, environments, tools, and MCP surfaces to one governed workspace.
Policy engine
Use a constrained YAML DSL that is readable by platform and security teams, then compile and publish policies per tenant.
Runtime decisions
Inspect context synchronously and evaluate actions before an agent can send mail, deploy, export, or mutate systems.
Approvals and incidents
Escalate high-risk actions to reviewers, persist the decision path, and triage incidents from the same timeline.
Run replay
Investigate artifacts, findings, action decisions, approvals, and outcomes without reconstructing the chain manually.
Commercial controls
Tie governed usage to plan limits for agents, runs, users, retention, Slack approvals, and export capabilities.
Credential broker
Store access once in encrypted form and let agents reach approved API targets only through short-lived broker leases.
Default Packs
Shipped policy packs for the first three agent classes
The v1 launch intentionally focuses on the agents that cause the fastest real-world harm if left ungoverned.
Browser agent
Escalate form submissions and high-risk web context before anything leaves the browser boundary.
Coding agent
Deny production deploys and gate risky repository or delivery actions.
Support agent
Block sensitive external email and customer-data leakage from conversational workflows.