Architecture decision guide

Choose the smallest AI system that can survive production

Framework rankings age quickly and hide the real tradeoff. Start with the workflow, authority boundary, failure modes, and operating cost. Choose tools only after those constraints are explicit.

Five tests before you compare frameworks

Start with the workflow

Write down the trigger, required context, permitted actions, accountable owner, and the point where a person must take over.

Separate retrieval from action

RAG answers from governed sources. Agents take actions through tools. Many systems need one of these, not both.

Model failure before framework choice

List unsafe outputs, unavailable dependencies, duplicate actions, timeouts, and recovery paths before selecting orchestration software.

Price the operating path

Include model usage, tool calls, evaluation, monitoring, review labor, incident response, and migration cost in the decision.

Demand a reversible first milestone

The first build should prove one workflow against acceptance tests without locking the team into an unnecessary platform.

Four operating patterns

Add autonomy only when the workflow requires it

Start at the left side of the complexity curve. Move to agents or orchestration only when a simpler pattern fails a written acceptance test.

01Low burden

Deterministic workflow

Choose when: Inputs, decisions, and permitted actions can be expressed as stable rules.
Avoid when: The work depends on interpretation across variable or incomplete context.
First proof: One idempotent path with retries, audit history, and a named exception owner.

02Medium burden

Retrieval assistant

Choose when: People need answers grounded in governed documents or structured knowledge.
Avoid when: The system must take consequential actions rather than provide cited answers.
First proof: A representative question set with citations, access controls, and answer evaluations.

03High burden

Tool-using agent

Choose when: Variable context requires bounded reasoning before selecting and using approved tools.
Avoid when: A workflow can encode the decision path without model judgment.
First proof: One bounded task with policy gates, approval points, failure recovery, and cost evidence.

04Very high burden

Multi-agent orchestration

Choose when: Distinct specialist roles or parallel review measurably improve a complex decision.
Avoid when: A simpler pattern—one agent or a deterministic workflow—meets the acceptance threshold.
First proof: A controlled comparison showing better accepted output than the simpler alternative.

$3K-$5K

Architecture and delivery sprint

Scope the system, map integration risk, and leave with an architecture, delivery plan, and build estimate.

Check architecture fit

Five tests before you compare frameworks

Start with the workflow

Write down the trigger, required context, permitted actions, accountable owner, and the point where a person must take over.

Separate retrieval from action

RAG answers from governed sources. Agents take actions through tools. Many systems need one of these, not both.

Model failure before framework choice

List unsafe outputs, unavailable dependencies, duplicate actions, timeouts, and recovery paths before selecting orchestration software.

Price the operating path

Include model usage, tool calls, evaluation, monitoring, review labor, incident response, and migration cost in the decision.

Demand a reversible first milestone

The first build should prove one workflow against acceptance tests without locking the team into an unnecessary platform.