Choose the smallest sufficient agent system

Agents are an ingredient of Technology capabilities, not a separate stage of the business decision loop.

The best agent system is not the one with the most autonomy. It is the least complex system that can complete the job within its authority, survive likely failure, and return evidence a person can use.

An agent system combines a model with some mix of goals, context, memory, tools, rules, and feedback. Add each layer only when the job requires it. Every extra layer creates another place for cost, latency, ambiguity, or unauthorised action to enter.

Gate: should this be an agent at all?

Use deterministic software when the inputs, rules, and expected output are known. Use a single model call when the job is one bounded transformation such as classifying, extracting, summarising, or drafting from supplied context.

Consider an agent only when the work must choose and sequence steps, inspect changing state, use tools, or recover from intermediate results.

Do not proceed until you can name:

• the outcome and beneficiary;
• the human who owns approval, accepted risk, and the stop decision;
• the data and tools the system may access;
• the actions it may and may not take;
• the evidence that will distinguish success from plausible-looking output.

If those boundaries are unclear, the next move is to define the job—not to add autonomy.

Check: what shape does the job have?

One bounded reasoning task

Choose a direct model call or a prompted assistant for explanation, drafting, classification, extraction, and other work that does not need to act across multiple systems.

Next move: give it representative inputs, an explicit output contract, and a small evaluation set. Keep the workflow at this level if it reliably passes.

One goal with a short tool path

Choose a single tool-using agent when the job needs a few adaptive steps, such as researching a question, updating a bounded record, or completing a defined workflow. Keep memory scoped to the task unless continuity demonstrably helps.

Next move: permit only the required tools, set a time or cost limit, name a stop condition, and review the first real run.

Repository-aware software work

Choose a coding-agent harness when success depends on repository instructions, code discovery, edits, tests, and version-control evidence. Codex, Claude, and Cursor harnesses are examples of this category; provider choice matters less than the surrounding permissions, instructions, checks, and review boundary.

Next move: start with one bounded change and require the relevant lint, type, test, and diff checks before choosing a specific harness.

A continuing thinking and operating partner

Choose a persistent assistant when work spans sessions and benefits from permissioned memory, recurring routines, or a durable decision history. Hermes is an example of this pattern: a continuing operating partner rather than an oracle or an owner of human purpose.

Next move: begin with one recurring decision loop. Define what may enter memory, what requires fresh consent, how evidence returns, and when stored material is corrected or removed.

A specialist interaction or infrastructure job

Choose a specialist system when the medium or trust requirement is the hard part of the job. Examples include conversational voice agents, browser-based operators, software-verification agents, and agents that need durable, independently verifiable storage.

Next move: test the specialist constraint directly—such as interruption handling for voice, reproducible findings for verification, or retrieval and integrity checks for durable memory. Do not let a specialist component become the system of record merely because it is convenient.

Several independent jobs with a real coordination need

Choose a multi-agent workflow only when the work separates into genuinely independent roles, benefits from parallelism or opposing perspectives, and has an explicit synthesis owner. A useful split is maker and checker; a weak split is several agents repeating the same vague request.

Next move: first run the workflow with one agent. Add a second role only when a measured bottleneck or verification gap justifies the coordination cost. Give every handoff an input, output, authority boundary, and acceptance check.

A long-running or public autonomous actor

Choose a persistent autonomous system only when unattended operation is itself necessary. Public communication, financial transactions, identity, secrets, and irreversible actions raise the proof burden sharply.

Next move: keep consequential actions behind human approval, constrain budgets and credentials, log attributable actions, rehearse recovery, and prove the stop mechanism before expanding exposure.

Check the fit before choosing a product

Compare candidate systems against the same use case. Ask:

• Can the simpler option complete the job?
• Does the system have only the authority it needs?
• Can it use the required tools without receiving unrelated access?
• Is memory necessary, permissioned, correctable, and portable?
• Are failures observable, bounded, and recoverable?
• Can an independent check distinguish success from confident error?
• Do the expected gains justify latency, operating cost, and maintenance?
• Can you change models or providers without losing the whole operating loop?

A product demo can show capability. It cannot prove fit for your environment. Treat each selection as a bounded experiment on representative work.

Next move: run the cheapest useful trial

1. Write one real task and its acceptance check.
2. Start with deterministic software or one model call.
3. Add one missing capability—tools, retrieval, memory, or coordination—only when the trial exposes its need.
4. Record quality, failures, human correction, time, and cost.
5. Continue, revise, or stop from that evidence.

Return to the Playbook when the outcome, authority, or protected constraint is still unclear.

When the smallest sufficient system is clear, return to Capabilities with the beneficiary, human owner, permission boundary, stop condition, and evidence check intact. An agent is one possible ingredient; continue without one when a person, deterministic tool, or simpler process is sufficient.

This routing is guidance, not a claim that one architecture is universally best. The right system remains conditional on the job, its environment, and the evidence it must return.