Skip to main content

Architecture decision guide

Pick the smallest system that survives production.

Framework rankings age fast and hide the real tradeoff. Start with your workflow, the authority boundary, how it fails, and what it costs to run. Pick tools after that.

Five tests before you compare frameworks

01

Start with the workflow, not the tool

Write down the trigger, the context it needs, what it may act on, who owns it, and the moment a person has to take over. The architecture falls out of that.

02

Decide whether you need answers or actions

RAG answers from your sources. Agents do things through tools. Most systems need one of those, not both — and building both costs you twice.

03

Work out how it breaks first

List the unsafe outputs, the dependency that goes down, the duplicate action, the timeout, and how you recover. Do it before you shortlist orchestration software.

04

Count what it costs to run

Model calls, tool calls, evals, monitoring, the person reviewing output, the migration later, and the incident you have not had yet.

05

Make the first build easy to walk away from

Prove one workflow. If that means you can leave the platform afterwards without rewriting everything, you picked well.

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.

Your next decision

Match the engagement to what is already known

An architecture sprint is the default when scope is still uncertain. Move directly to a build or rescue only when the evidence supports it.

The workflow, integrations, or acceptance tests are unresolved

Architecture and delivery sprint

$5K-$8K

Map the Solana or AI system boundary, authorities, integrations, risks, KPIs, and acceptance tests before committing to a build.

Send architecture brief

The workflow is validated and an owner can accept the build

Production build

$10K-$40K

Ship a Solana program or product surface first; also operational AI, game systems, and integrated business software against written acceptance criteria.

Send production brief

An existing system is failing, drifting, or blocking launch

Reliability and rescue sprint

$5K-$15K

Stabilize live Solana programs, agent workflows, or product systems—tests, observability, and the highest-risk paths first.

Send rescue brief