AI Agent Architecture Designer

Runs in your browser

Use this when: defining an agent's tools, permissions, approval gates and failure handling before development—not just sketching its purpose.

WHAT YOU'LL ACHIEVE

Create a human-reviewed architecture blueprint for a proposed agent, including tasks, permissions, approvals, safeguards and acceptance criteria.

THE TOOL: ARCHITECTURE BLUEPRINT

Make the decisions visible.

This is a design aid, not a working agent. Examples are prompts only; no tools, integrations or data sources are connected.

0 of 13 required design decisions recorded. Still decide: Agent purpose; Users and affected people; In-scope and out-of-scope boundaries; Tasks and hand-offs; Proposed autonomy level; Tools and integrations (planning only); Data sources and ownership; Data sensitivity; Permissions and least-privilege boundary; Human-in-the-loop approval gates; Errors, fallback and escalation; Safety and prompt-injection mitigations; Acceptance criteria and decision owner.

SECTION 1 OF 5

Purpose, users and scope

Start with a bounded problem and the people affected.

Agent purposeWhat outcome should this proposed agent help achieve?Required decision.Users and affected peopleWho operates it, relies on its output or may be affected?Required decision.In-scope and out-of-scope boundariesName forbidden or deferred uses, not just the intended use.Required decision.

SECTION 2 OF 5

Tasks, autonomy and planned tools

Describe a proposal, not a live integration or deployed agent.

Tasks and hand-offsList the steps, inputs and expected outputs; identify where a person takes over.Required decision.Proposed autonomy levelChoose the strongest autonomy this design would permit; it is not enabled here.Required decision.Tools and integrations (planning only)Name proposed capabilities and actions. Nothing is connected by this planner.Required decision.

SECTION 3 OF 5

Information and permissions

Limit access to the minimum needed for the defined task.

Data sources and ownershipRecord source owner, purpose, access route and any exclusions.Required decision.Data sensitivityUse the highest sensitivity that may be encountered; resolve unknowns before design approval.Required decision.Permissions and least-privilege boundarySpecify identity, read/write limits, secret handling and how access is revoked.Required decision.

SECTION 4 OF 5

Human control, memory and failure handling

Make approval points and recovery behaviour explicit before considering implementation.

Human-in-the-loop approval gatesSay who reviews what, when approval is mandatory and how a person can stop or override.Required decision.Memory and state assumptionsDescribe what may persist, for how long, who can inspect it and how it is cleared. Unknown means no assumption of safe persistence.Errors, fallback and escalationCover unavailable tools, ambiguous requests, low confidence, timeouts and harmful or conflicting outputs.Required decision.

SECTION 5 OF 5

Safety, evaluation and acceptance

Define protections and evidence a human owner would require before any later implementation.

Safety and prompt-injection mitigationsConsider untrusted content, instruction hierarchy, tool allow-lists, output checks and data exfiltration.Required decision.Monitoring and evaluation planName observable signals, representative tests, owner, review cadence and incident path.Acceptance criteria and decision ownerState measurable conditions and who can accept, reject or defer the proposed design.Required decision.

HUMAN-READABLE SUMMARY

Review the proposed design.

Decisions still needed

Agent purpose
Users and affected people
In-scope and out-of-scope boundaries
Tasks and hand-offs
Proposed autonomy level
Tools and integrations (planning only)
Data sources and ownership
Data sensitivity
Permissions and least-privilege boundary
Human-in-the-loop approval gates
Errors, fallback and escalation
Safety and prompt-injection mitigations
Acceptance criteria and decision owner

Show proposed flow and decision points

This optional view is a planning aid, not executable logic.

Confirm purpose, users and scope.
Check that the requested task is in bounds.
Use only proposed, least-privilege tools and approved sources.
Require the stated human approval before consequential action.
Stop, explain and escalate on error, uncertainty or denied access.
Review monitoring evidence against acceptance criteria.

AI AGENT ARCHITECTURE DESIGN AID

PURPOSE, USERS AND SCOPE
Agent purpose: Not decided
Users and affected people: Not decided
In-scope and out-of-scope boundaries: Not decided

TASKS, AUTONOMY AND PLANNED TOOLS
Tasks and hand-offs: Not decided
Proposed autonomy level: Not decided
Tools and integrations (planning only): Not decided

INFORMATION AND PERMISSIONS
Data sources and ownership: Not decided
Data sensitivity: Not decided
Permissions and least-privilege boundary: Not decided

HUMAN CONTROL, MEMORY AND FAILURE HANDLING
Human-in-the-loop approval gates: Not decided
Memory and state assumptions: Not decided
Errors, fallback and escalation: Not decided

SAFETY, EVALUATION AND ACCEPTANCE
Safety and prompt-injection mitigations: Not decided
Monitoring and evaluation plan: Not decided
Acceptance criteria and decision owner: Not decided

DESIGN FLOW (PROPOSED ONLY)
Defined purpose and users
→ Check scope and permitted tasks
→ Use planned tools only: Not decided
→ Apply access and data boundaries: Not decided
→ Pause at human approval gates: Not decided
→ On failure, use fallback or escalation: Not decided
→ Evaluate evidence against acceptance criteria before any separate implementation decision

Status: This is a human-readable planning aid, not a working or deployed agent. Tool and integration entries are proposals only. No services are connected and no data is accessed. Incomplete decisions: Agent purpose; Users and affected people; In-scope and out-of-scope boundaries; Tasks and hand-offs; Proposed autonomy level; Tools and integrations (planning only); Data sources and ownership; Data sensitivity; Permissions and least-privilege boundary; Human-in-the-loop approval gates; Errors, fallback and escalation; Safety and prompt-injection mitigations; Acceptance criteria and decision owner. Inputs remain in this page session only.