Agentic AI System Design / Canvas

The Agentic System Design Canvas

Twelve sections that force the decisions an agent design usually skips: what the model may never decide, which effects cannot be repeated, what survives a crash, and which number blocks a release. Fill it for Atlas, for a practice scenario, or for the system you are actually being asked to build at work.

HOW TO USE IT

Work top to bottom the first time. Sections 1 to 3 determine everything after them: if the outcome and the risk tier are vague, the topology and the controls will be too. Leave a field blank rather than filling it with a placeholder, because the blanks are the honest map of what you still have to find out.

UNDER TIME PRESSURE

Outcome and risk
Authority boundaries
Topology
Tools and context
State and durable execution
Security and human review
Evals and SLOs
Scale, cost, and rollout

This canvas saves to your browser's local storage as you type. Nothing is uploaded, nothing is transmitted, and no account exists. Clearing site data clears the canvas, so export anything you want to keep.

DESIGN CANVAS

Atlas capstone

Twelve sections, 72 fields. Everything you type is saved in this browser as you go. Nothing is uploaded, and nobody else can read it.

0 / 72 filled

01 Outcome

Start with the job, not the model. If you cannot name the metric, you cannot design the system.

User

Job to be done

Business outcome

Success metric

Non-goals

02 Why agentic?

Agentic is a cost, not a default. Justify the judgment you are buying.

Why a deterministic workflow is insufficient

Where model judgment is useful

Where deterministic code remains authoritative

Simpler alternative considered

03 Risk and authority

Name what the model may never decide. The risk tier shapes the architecture; it is not a paragraph added at the end.

Risk tier 0 read-only, 1 reversible, 2 material effect, 3 regulated, 4 not automatable

Worst credible failure

Read-only actions

Reversible writes

Irreversible or material actions

Human approval points

Prohibited autonomous decisions

04 Journey and topology

The simplest topology that satisfies the requirements, plus the conditions that stop it.

Happy path

Agent or workflow pattern workflow, tool loop, router, planner-executor, supervisor, durable workflow

Key states

Stop condition

Escalation condition

Cancellation behavior

05 Tools

Every tool is a contract: typed, scoped, rated for side effect, and idempotent where it touches the world.

06 Context and data

Rank sources by authority and freshness, and keep untrusted content on the far side of a boundary.

Trusted instructions

User inputs

Retrieved sources

Authority and freshness ranking

ACL strategy

Citation and provenance

Context budget and compression

Untrusted-content boundary

07 State, memory, and durability

Say what survives a crash and what must never happen twice.

Request state

Run state

Thread or session state

Workflow history

Long-term memory

System of record

What survives a crash

What must never repeat

Resume and version strategy

08 Security and governance

Identity has to reach the data and the tools. Injection controls are architecture, not prompt text.

Trust boundaries

Tenant isolation

Secrets and service identity

Prompt-injection controls

Data-exfiltration controls

Sandbox and egress controls

Retention and deletion

Audit events

09 Evals

Measure the trajectory, not only the final text, and name the number that blocks a release.

Golden cases

Adversarial cases

Component metrics

Trajectory metrics

Outcome metrics

Safety and policy metrics

Human calibration

Release gate

10 Operations

One quantitative objective per axis, and the person who is paged when it breaks.

Latency SLO

Availability and durability SLO

Cost budget

Throughput and peak load

Rate-limit and backpressure strategy

Critical alerts

Runbook owner

11 Rollout

How the change reaches production, and what pulls it back.

Offline test

Shadow

Canary

Rollback trigger

Versions recorded

Production sampling

12 Open tradeoffs

Three decisions you made knowing what they cost. A design with no tradeoffs is a design nobody stress-tested.

Tradeoff 1

Tradeoff 2

Tradeoff 3

Per-scenario copies

Each practice scenario keeps its own canvas, so a design for one system never overwrites another. Open a scenario and expand the canvas at the bottom of the page.