How Many Credentials Should Your AI Agent Have? Zero. — Jim Clark, Docker
- Source
- youtube.com
- Author
- AI Engineer
- Date
Context
Jim Clark, a principal software engineer at Docker, argues in an AI Engineer talk that an 's safety depends less on its control loop than on which tools and reach it, so the unit to manage is the , an isolated environment where it runs. The talk's published description gives two examples: a newsroom split into researcher, fact-checker and publisher sandboxes, so untrusted input is not mixed with dangerous tools, and a coding agent that receives signing keys only while committing.
Clark's rule is that a sandbox should hold zero credentials. Each sandbox connects through one MCP gateway, a single point where requests for tools can be allowed or refused (MCP is a standard way for agents to call tools), and Okta's Cross App Access lets agents use a company's existing single sign-on. The description does not explain how access is granted when a secret is needed, and only the description was reviewed.
- Clark argues that an is a simple loop, so safety comes from controlling which context and tools flow into it. That makes the sandbox, not the harness, the boundary to manage.
- His newsroom example splits one workflow into separate researcher, fact-checker and publisher sandboxes, and his coding-agent example grants signing keys only while the agent is committing.
- Giving each sandbox a single MCP gateway endpoint creates one control point for tools, resources and prompts, and Clark says it also makes harnesses interchangeable.
- Cross App Access (XAA) with Okta ties agent identity and authorization grants to an organization's existing SSO, which supports Clark's position that the right number of credentials stored in a sandbox is zero.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
- sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
- agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
Checking sign-in…
Loading comments…






