Blog
From the Dome team
News, practical guidance, and analysis for teams operating AI agents.
All posts
Investigate an MCP tool call without mistaking permission for success
Follow a hypothetical ticket-close request from permission to backend outcome, and decide what to check before retrying it.
Guide · Tu Nguyen
Prove an audit export is complete for its stated scope
Deliver an audit export with a manifest covering scope, time windows, limits, pagination, unresolved records, and collection gaps.
Guide · Preska Sharma
Delegated agents: decide what happens when the user is missing
Decide whether queued work still needs a verified user or has been explicitly accepted under a separately scoped service identity.
Guide · Tu Nguyen
Govern MCP access with a discovery-and-invocation test matrix
Test tool discovery, invocation, backend ownership, and alternate routes separately before approving an MCP integration.
Guide · Preska Sharma
Choose a filter or validator for a tool content contract
Use a tool response contract to choose between removing sensitive content and rejecting missing or invalid structure.
Explainer · Preska Sharma
Roll out default-deny without learning permissions from outages
Roll out default-deny with explicit allowed and forbidden cases, a bounded first workload, and a tested rollback plan.
Guide · Tu Nguyen
Retire an agent without leaving its authority behind
Check owners, scheduled work, descendants, and separate credentials before permanently revoking an agent and recording what remains to clean up.
Guide · Preska Sharma
What should an agent’s budget buy?
Use a declared outcome, cost basis, and retry record to judge whether an agent budget is buying useful completed work.
Analysis · Tu Nguyen
Handle signed Dome webhooks without repeating the business action
Design a webhook receiver that verifies deliveries, accepts work durably, and prevents retries or replay from opening duplicate review tickets.
Guide · Preska Sharma
When a model pool should fail over, skip a candidate, or stop
Use a routing decision table to distinguish provider failures from policy refusals, quota exhaustion, and interrupted model streams.
Guide · Tu Nguyen
Return only the tool fields an agent needs
Use a synthetic directory response to test that a filter removes private fields while preserving the data the agent needs to finish its task.
Guide · Preska Sharma
Should the same refund request get three different answers?
Use a synthetic refund fixture to test whether verified delegated identity and authoritative resource facts produce the required authorization outcome.
Guide · Tu Nguyen
Test agent authorization beyond the requests in your logs
Build a policy test matrix for allowed, denied, and missing-context requests, then review it alongside historical decision differences.
Guide · Tu Nguyen
Trace an authorization decision across inherited rules
Find which rule or managed setting caused a surprising decision, then test the change across organization, tenant, workspace, and agent scopes.
Guide · Preska Sharma
ABAC vs. RBAC for agents: choose the facts your policy needs
Work through a deployment example to choose roles, attributes, and backend checks based on the facts that should change an access decision.
Explainer · Tu Nguyen
Choose a standing or delegated agent
Decide whether a task runs under a service’s own authority or represents a verified person, including when queued work executes later.
Explainer · Preska Sharma
Design a model pool as an application contract
Define what every model in a pool must preserve: approved data handling, usable output, routing preferences, and a tested change procedure.
Explainer · Tu Nguyen
What an agent audit trail needs to show
Use an investigation worksheet to connect an agent request to its identity, authorization decision, execution record, and backend outcome.
Explainer · Preska Sharma
Build an agent inventory you can operate
Compare deployments, registered identities, owners, and recent requests to find the gaps another on-call operator would need to resolve.
Guide · Tu Nguyen
Choose rules, guards, or quotas for the job
Decide whether a requirement needs an authorization rule, a content guard, or a usage quota, then test each control against a ticket-read example.
Explainer · Preska Sharma
When an agent acts for you, whose permissions should count?
An agent and the person it represents have different permissions. Use a refund example to decide which identities, entitlements, and transaction facts to check.
Analysis · Tu Nguyen
Move tool credentials out of agents without moving the risk out of sight
Move backend secrets out of agent deployments with a staged cutover, a bounded rollback plan, and separate recovery steps for agent and upstream credentials.
Guide · Preska Sharma
MCP authentication vs. authorization: what a token allows
Follow an incident-management example to see what OAuth grants, what a valid token proves, and where permissions for individual actions are checked.
Explainer · Tu Nguyen
Review a remote MCP server before connecting it
Work through a support-ticket example to check an MCP server's permissions, credentials, data handling, and failure behavior before connecting an agent.
Guide · Tu Nguyen
Dome is generally available, and now self-serve
Three things became clear over months with dozens of organizations, and all of them point at the agentic loop rather than its endpoints. Here is what you can run today.
News · Dave McJannet
Dome Systems launches access to its Agent Operations Platform
Platform teams can now sign up and start governing their AI agents: an agent registry, an MCP gateway and LLM model routing, on one composable set of controls.
News · Dome Systems