AnthropicClaude Certified Architect – Professional2026-103.1 Evaluate tool/agent configuration for capability bloat
Remove privilege without breaking the workflow
Completed by the knowledge check
Learning objective
Apply least privilege by removing capabilities a role does not need and splitting broad agents into scoped specialists without fragmenting critical workflows.
Core idea
An explicit role menu and a constrained tool implementation establish the capability boundary. The architect’s next decision is what to remove, where retained authority belongs, and whether a specialist split preserves the work the business actually needs.
Remove unnecessary authority before adding controls around it. Logging can expose misuse, and confirmation can constrain a retained action, but neither explains why a document reader should hold a write capability. Conversely, splitting every operation into a separate agent can turn one coherent transaction into a chain of incomplete handoffs. Least privilege must preserve a complete, authorized route to the required outcome.
How it works
Turn audit findings into removal decisions
Use the dispositions from 3.1.1 to produce an explicit role-to-capability configuration. For a Messages application, the enforced artifact is the tools array the trusted request builder supplies. Exclude unneeded definitions there. Derive role selection from the application’s verified caller context rather than a sentence in the conversation claiming a privileged job.
Make unknown roles resolve to an application-defined no-tool path or a rejected request. That default is your request-builder policy, not a special Anthropic role field. Assign ownership of the builder and its role table to the platform team, while business operation owners approve the required capabilities.
For tools exposed through the Messages API MCP connector, set the mcp_toolset entry’s default_config to { "enabled": false } and enable only required names through configs: { "REQUIRED_TOOL": { "enabled": true } }, replacing REQUIRED_TOOL with the server’s tool name. Because enabled defaults to true, an unconfigured toolset exposes every tool the server offers. The connector is beta on the Claude API, Claude Platform on AWS, and Microsoft Foundry, requires anthropic-beta: mcp-client-2025-11-20, and is not eligible for zero data retention (ZDR).
Check alternate routes before accepting the cut. If a removed write remains possible through an unrestricted administrative tool or a callable specialist, the workflow still holds that authority. Follow the implementation to the side effect, then remove or narrow every route the role does not need. Backend resource authorization remains necessary for retained operations; the role menu does not establish which individual record the caller may access.
The workflow still holds a removed write while an unrestricted administrative tool or a callable specialist reaches its side effect.
Choose removal rather than approval settings
Recall that bare-name disallowedTools (Python: disallowed_tools) removes a definition, while pre-approval and permissionMode: "dontAsk" (Python: permission_mode="dontAsk") do not.
These distinctions matter when defending a production decision. “The role cannot receive the tool definition” is a narrower, inspectable claim than “the role normally asks before using it.” Keep confirmation for high-impact actions the role genuinely needs, and keep an execution record for those actions. Neither is the primary remedy for surplus authority.
Split around responsibility and complete outcomes
Give a specialist the smallest coherent assignment that can produce a useful result. A read-only case reviewer and a separately authorized scheduler can be sensible roles because their authority differs. A different agent for each field lookup usually adds handoffs without establishing a meaningful authority boundary.
Implement the split through explicit AgentDefinition.tools menus; the broader delegation design belongs to 1.4.3.
Review dependencies before drawing the boundary. Which evidence does the writer require? Who accepts a missing fact? Which component owns the final transaction? Anthropic’s tool-design guidance supports consolidating related operations into a useful workflow tool rather than wrapping every low-level endpoint. Preserve that advantage where the same authority applies, instead of requiring separate agents to coordinate each API call.
Use an application-owned narrow operation when a specialist needs a small piece of another domain. For example, a case reviewer can receive an approved appointment-availability lookup without inheriting all scheduling administration. Do not restore a general-purpose tool to repair one missing input.
Price the split and assign its recovery owner
Anthropic documents that a subagent’s intermediate calls and results stay in its own conversation; the parent receives its final message. That isolates working context, but creates a handoff obligation. Specify what the parent must receive to continue, and what it does when the specialist returns an unresolved dependency instead of a completed result.
Compare complete-workflow success and time with the unsplit configuration. If the narrower specialist forces the coordinator to repeat investigations or loses a required identifier, repair the handoff or change the boundary. Do not count a successful local subtask as successful business delivery. Use 3.1.4 to validate the whole redesigned route, including rejection and recovery.
Key distinctions and trade-offs
Removal versus monitoring: Removal eliminates that exposed capability from the role. Monitoring records activity on the capabilities that remain. Retain both where appropriate, with different acceptance criteria.
Specialist isolation versus workflow fragmentation: Separate authority and independent work justify a boundary. Repeated transfers of the same transaction state are a cost that the architect must measure and accept.
Narrow tool versus generic convenience: A coherent workflow tool can reduce calls without broadening privilege. A generic administrative wrapper reduces names while expanding possible actions. Judge the reachable operations, not the visual size of the menu.
Architecture / decision scenario
A healthcare provider’s referral service must prepare a referral and offer an appointment within 90 seconds. Its intake agent can read submitted records, cancel appointments, edit clinical notes, and change scheduling rules. Intake needs none of the latter three operations. All tool names in this scenario are application-defined.
The service owner proposes keeping the broad agent and asking for confirmation on every write. The architect removes the unneeded operations from intake’s Messages tools array instead. A referral reviewer prepares the required facts; a scheduling specialist receives only the appointment operations needed for this route. The scheduling implementation still authorizes the selected patient’s appointment.
A proposed four-agent split assigns one agent to each lookup. In a hypothetical pilot it raises complete-workflow time from 62 to 104 seconds, exceeding the target. The architect keeps related referral lookups together behind a narrow workflow operation and preserves the authority boundary around scheduling. The clinical operations owner accepts referral completeness; the scheduling owner accepts the authorized booking path.
A missing appointment identifier produces an unresolved handoff, not permission for intake to regain scheduling administration. The coordinator requests the missing lookup through the retained narrow route. This design removes surplus writes while preserving a complete path to a useful appointment offer.
Common exam traps
- Adding confirmation to an operation the role never needs.
- Leaving an MCP toolset unconfigured and exposing every operation from its server.
- Accepting a smaller direct menu without tracing the authority retained through specialists.
- Removing a direct operation while leaving an equivalent delegate reachable.
- Splitting by low-level endpoint and ignoring transaction completion time.
- Restoring an entire administrative menu to repair one missing lookup.
What to remember
Enforce removal where the role’s capabilities are assembled. Split around meaningful authority and complete responsibilities, then measure the full workflow. A narrower configuration succeeds when prohibited actions are absent and legitimate work still reaches an authorized outcome.
Knowledge check
A logistics company connects its dispatch assistant to a carrier’s MCP server through the Messages API MCP connector. The server offers 40 tools. The dispatch role needs only get_shipment and track_parcel, and must not be offered any other tool, including tools the carrier adds later. How should the mcp_toolset entry be configured?
get_shipment and track_parcel in configsdefer_loading: true for all tools in the toolset’s default_configdefault_config and enable the two in configsconfigs empty and name the two tools in the system promptenabled defaults to true , so a toolset that only enables two tools in configs still exposes the other 38, and any tool the carrier adds later. Setting default_config to { "enabled": false } turns the default into an allowlist, and configs then enables only the two required names. Deferral changes when definitions load, not which tools the role can call.An Agent SDK agent summarizes contracts and must never write files. The security review requires that the agent not receive the Write tool definition at all. Which configuration meets that requirement?
Write to disallowedToolspermissionMode to "dontAsk" for the sessionWrite out of the allowedTools listEdit(/contracts/**) to disallowedToolsWrite . dontAsk denies calls that would prompt, but the definition stays in context. A tool missing from allowedTools is still available and falls through to the permission mode. A scoped Edit rule restricts writes matching its path and leaves the Write definition in place.An insurer splits its claims-intake agent so that intake no longer holds payment authority. The proposed design gives each lookup — policy, coverage, prior claims, adjuster notes — its own read-only agent, plus a separately authorized payments specialist. All four lookups use the same read authority. In a pilot, complete-claim time rises from 70 to 125 seconds against a 100-second target. What should the architect change?
After a redesign, a hospital’s discharge reviewer is read-only, and only a scheduling specialist can book appointments. Reviews now stall because the reviewer cannot see whether a follow-up slot exists before it recommends a discharge date. What should the architect add?
References
- Anthropic, MCP connector
- Anthropic, Define tools
- Anthropic, Configure permissions
- Anthropic, Subagents in the SDK
- Anthropic, Writing effective tools for agents
- Anthropic, Mitigate jailbreaks and prompt injections