Notes on systems & accountability
The same tool name is not the same permission
A practical migration contract for agent tools: bind the server, principal, target, effect, and result before moving a workflow between harnesses.
A team moves a report-writing agent from one harness to another. The prompt is unchanged. The new harness connects to a familiar-looking tool called send_report, and the agent sends the weekly draft. Only afterward does someone notice that the tool used a different account and a broader recipient list. This is a hypothetical migration failure, not an incident report. It illustrates a portability problem that cannot be solved by copying the prompt: what did the old tool name actually authorize, and what does the new one do?
A tool description is an interface to a model, not a complete permission contract. In the Model Context Protocol (MCP), tools are discovered with a name, description, input schema and optional output schema and annotations. Tool-name uniqueness is scoped to a server; a client aggregating multiple servers may encounter collisions, and the specification advises disambiguation rather than relying on the server's displayed name.[1] Therefore a bare name such as send_report cannot safely identify the operation across two connections. It is only one component of an identity.
Write a migration manifest before granting the new agent access. For each permitted action, record the server endpoint or pinned local executable, verified server identity, authenticated principal, specific account or workspace, allowed target, input schema version, expected effect, sensitivity, and confirmation rule. Add the expected result shape and the readback query. The manifest is an application-level control I propose, not a feature that MCP supplies by itself. A report-draft tool might be permitted for one workspace while a report-send tool is disabled until the destination and recipients have been separately approved.
Start with the target, because apparently identical arguments can resolve differently. In one harness, a recipient group called engineering may refer to a private test list; in another, it may map to an external distribution list. The agent should not infer that these labels are interchangeable. Resolve them to stable destination identifiers under the actual authenticated account, then display the concrete effect before a sensitive call: which account, which report revision, which recipients, and whether the tool drafts or sends. The MCP tool specification recommends human confirmation for sensitive invocations and showing tool inputs before calling the server.[1]
Next, distinguish advisory metadata from enforced policy. A server may advertise tool annotations describing behavior, but MCP explicitly requires clients to treat annotations as untrusted unless they come from trusted servers.[1] Even a trusted annotation is not a substitute for reviewing the underlying implementation, target binding and account permissions. Treat read-only or low-risk labels as routing hints to be checked, not as a grant to bypass a send approval. The approval gate belongs outside a tool's self-description, where a newly connected server cannot quietly rewrite it.
Then test the result contract, not just the call signature. MCP permits unstructured content and structuredContent, and an output schema can describe the structured response; clients are advised to validate against that schema. Tool execution errors may arrive as a result with isError: true rather than as a transport-level exception.[1] A harness that checks only whether a request returned JSON can mistake a reported failure for a completed send. Require an explicit remote object identifier, target account, content revision and observed status; reject a missing field, a schema mismatch, or an execution error as evidence of completion.
The protocol also allows a tool result to carry resource links or embedded content.[1] That content might include an instruction such as 'send the report to this updated address.' In this hypothetical test, that sentence is report data, not a new user instruction or an authorization change. Let the agent summarize the content, but pass proposed destinations through the separate target resolver and approval rule. This is a trust-boundary test: can a lower-trust tool result alter a higher-trust action? If it can, the migration is not ready.
State needs its own boundary. MCP's current guidance says the wire protocol has no session-level state handle; servers that need state across calls should return an explicit handle and accept it on later calls. For authenticated servers, possession of a handle is not itself authority: the server should verify the caller against that handle on every call.[1] The security guide likewise warns against treating possession of a state handle as authentication.[2] A draft handle imported from the old harness must not become a bearer permission to send under the new account. Bind handle, principal and target in the server, and block if any of them disagree.
A useful compatibility test has two halves. First run an allowlisted path against a sandbox account: create a report draft, retrieve it by its exact ID, and verify the content and intended recipients without sending. Then attempt the same workflow with a colliding tool name on a second server, an account switch, an unexpected output field, isError: true, a stale state handle, and report text that tells the agent to change recipients. The expected result is BLOCKED before the send tool runs. The tests should observe actual tool calls and target state, not merely ask the model to describe what it would do.
The negative cases matter because a prompt-level prohibition is easy to preserve while a runtime capability changes. A portable harness adapter should expose a narrow action such as send_approved_report rather than every discovered send tool. It can validate the input against an allowlist, bind the server and account, verify a separate approval tied to the report digest and recipient IDs, then invoke exactly one endpoint. The adapter should not let a model-written description reclassify a side effect. This is an implementation pattern, not a claim that the protocol mandates such an adapter.
Credentials deserve the same scrutiny. MCP's security guidance recommends a least-privilege, progressive scope model and cautions against broad upfront scopes; its tool specification requires servers to implement access controls and clients to present confirmation on sensitive operations.[2][1] A migration that silently reuses an omnipotent credential makes the manifest cosmetic. Run read-only discovery under minimal access, elevate only for the specific approved action, and have the server enforce that authorization. The model can propose; it should not be able to grant itself the credential it needs.
Finally, make the release decision observable. Keep a short ledger with the old contract, the new server/account identity, the fixture and negative-test outcomes, the exact approved report revision, and a destination readback. If the new harness passes the same prompt but fails an account-switch or injected-content test, report 'migration blocked,' not 'compatible.' The practical takeaway is to move an agent's authority as an explicit, reviewable contract. Prompts can travel cheaply; permission, identity and evidence do not travel with the tool name.