Review guide
How to apply Blueprint principles to a real feature or agent design before shipping.
The three-question review
Before shipping any agentic feature, ask: (1) Is every destructive or externally visible action gated by explicit user approval or a bounded autonomy model? (2) Can a user see what the agent is doing at any point without reading logs? (3) If the agent produces a wrong output, is there a path to undo, retry, or repair without data loss?
Cluster checklist
Delegation: have you defined what the agent can decide autonomously? Visibility: does every async action show visible progress? Trust: are risky actions gated? Orchestration: does the pipeline degrade gracefully when a subagent fails? If you cannot answer yes to all four, the feature is not ready to ship.
Using the MCP audit tool
validate_agent_architecture (Pro/Teams) accepts your implementation description and returns a structured review aligned to the four clusters. Pass the full context — system prompt, autonomy scope, approval flow, and output types. Set private_session=true to skip server-side logging. This replaces the manual checklist for production reviews.
Review checklist — all principles
For each principle: does your feature respect it? Flag the ones that need attention.
Design for delegation rather than direct manipulation
Design experiences around the assignment of work, the expression of intent, the setting of constraints, and the review of results, rather than requiring users to execute each step manually.
Review question: The product remains structurally manual, with AI functioning only as an overlay rather than as an operational capability.
Ensure that background work remains perceptible
When the system is operating asynchronously or outside the user’s immediate focus, it should provide persistent and proportionate signals that work is continuing.
Review question: The user cannot determine whether the task is active, delayed, blocked, or failed.
Align feedback with the user’s level of attention
The system should calibrate the depth and frequency of feedback according to whether the user is actively engaged, passively monitoring, or temporarily absent.
Review question: The system becomes either excessively opaque or unnecessarily burdensome.
Apply progressive disclosure to system agency
Provide the minimum information necessary by default, while enabling users to inspect additional detail when confidence, understanding, or intervention is required.
Review question: The system is experienced either as an opaque black box or as an over-exposed technical log.
Replace implied magic with clear mental models
The product should help users understand what the system can do, what it is currently doing, what it cannot do, and what conditions govern its behaviour.
Review question: Users attribute a level of agency, reliability, or competence that the system does not in fact possess.
Expose meaningful operational state, not internal complexity
Present the state of the system in language and structures that are relevant to the user, rather than exposing low-level internals that do not support action or understanding.
Review question: The system either obscures important status or overwhelms the user with irrelevant mechanics.
Establish trust through inspectability
Users should be able to examine how a result was produced when confidence, accountability, or decision quality is important.
Review question: The system demands trust without providing a basis for it.
Make hand-offs, approvals, and blockers explicit
When the system cannot proceed, the reason should be immediately visible, along with any action required from the user or another dependency.
Review question: Tasks stall silently, creating confusion and eroding confidence.
Represent delegated work as a system, not merely as a conversation
Where work involves multiple steps, agents, dependencies, or concurrent activities, it should be represented as a structured system rather than solely as a message stream.
Review question: Complex work is reduced to an unstructured narrative and becomes difficult to govern.
Optimise for steering, not only initiating
The system should support users not only in starting tasks, but also in guiding, refining, reprioritising, and correcting work while it is underway.
Review question: The only available control becomes repetition or restart, which is inefficient and often operationally unsafe.
Also in this section