Patterns for Real Systems

Separate understanding from implementation when the work is consequential or broad. The planning stage researches the environment, resolves assumptions, chooses an approach, and writes acceptance criteria. The build stage executes that bounded plan. The review stage compares the artifact and behavior against the plan and the real user outcome.

Review upstream thinking more intensely than generated volume. A flawed plan can produce thousands of internally consistent lines that solve the wrong problem. Check research, domain assumptions, architecture, scope, risk, and proof before implementation makes them expensive.

In an observe-and-act loop, the worker chooses one bounded action, receives an environmental result, updates task state, and decides what to do next. This is useful when the path cannot be known in advance: researching an unfamiliar issue, diagnosing a failing system, or navigating a changing external interface.

The loop needs stopping conditions, action budgets, error categories, and protection against repeating the same failed move. Tool output is evidence, not a command. Persist enough state that a retry or human handoff does not restart the investigation from zero.

An orchestrator retains the objective and delegates bounded work to specialists. It is valuable when tasks require different tools or forms of expertise but one owner must reconcile the results. Specialists can run sequentially, in parallel where independent, or on demand as tools.

The orchestrator should not merely relay prose. Give each specialist a typed input and return artifact, establish context boundaries, and define who integrates conflicts. Use a direct function or one worker when specialization does not improve quality enough to repay coordination cost.

A repeat-until-done loop can help when progress is objectively measurable: fix failing tests, process a known queue, reconcile records, or search until specified evidence is found. The loop should carry a stable objective, current state, remaining work, budget, and a validator that can end it.

Never use ‘keep going until it works’ as the only control. Add maximum turns, cost and time limits, duplicate-action detection, escalating retry policy, a no-progress condition, and a safe terminal state. Autonomy is a loop with rails, not permission to wander.

Do not load every instruction, tool, and document at startup. Begin with the task and a small capability index. Reveal specialized guidance, examples, schemas, or tools only when the current state requires them. This preserves attention and reduces accidental authority.

Recurring expert work can become a versioned skill: instructions, references, templates, scripts, and evaluation cases packaged together. Let corrections improve the skill, but require review before a worker rewrites the rules that govern its own behavior.

A person can participate inside the workflow by approving a specific action, resolving ambiguity, or handling an exception. A person can also supervise the loop from above by setting policy, reviewing metrics, examining failures, and deciding which action classes have earned more autonomy.

Do not convert human accountability into rubber-stamp clicking. Present the smallest meaningful decision with relevant evidence and recovery. Improve the workflow when approvals are routinely accepted, routinely revised, or asking the wrong person.

An expert swarm is appropriate when several independent analyses or work packages can create real coverage, disagreement is useful, and synthesis has an owner. Examples include research across distinct domains, parallel evaluation of alternatives, or implementation across isolated packages.

Production collaboration needs a task graph, ownership, shared identifiers, dependency state, budgets, cancellation, conflict policy, and integration verification. More agents increase the number of interfaces and possible partial failures. Treat swarm size as a cost to justify, not a maturity badge.

Use deterministic code for stable rules, a single model call for bounded semantic transformation, an observe-and-act loop for adaptive exploration, an orchestrator for genuine specialization, and a human gate where consequences require accountability. Combine patterns only when each boundary remains legible.

The best architecture often begins as a simple workflow and earns agentic freedom at one uncertain seam. Start where the environment can supply feedback and the outcome can be evaluated. Complexity should follow measured failure, not anticipation of scale.