⟶ Insights / AI Engineering
AI agent architectures, for real work.
The patterns senior teams actually ship - single-tool, ReAct, planner-executor, multi-agent - with the guardrails, evals and cost discipline enterprise workloads need. No frameworks-worship. No demoware.
11 min read · Updated Jul 2026
The short answer
Pick the simplest agent pattern that meets the SLA. Escalate one tier when - and only when - the previous tier's failure mode is documented, not suspected.
Most 'agent' failures in production are guardrail failures, not model failures. Max-step limits, tool allowlists, output validation and prompt-injection defence are the actual product. The LLM is a component.
The five patterns.
| Pattern | Shape | Where it fits |
|---|---|---|
| Single-tool assistant | One LLM call, one function. | Invoice extraction, structured Q&A, form-filling. |
| ReAct loop | Reason → act → observe, repeat until goal. | Research assistants, deep-lookup copilots, multi-step data pulls. |
| Planner-executor | One model plans; another executes each step. | Durable workflows, back-office automation, claims triage. |
| Multi-agent supervisor | Router delegates to specialised agents. | Cross-domain assistants, complex operational workflows. |
| Human-in-the-loop | Agent proposes; human approves before action. | Regulated, high-stakes or reversibility-limited steps. |
Frequently asked.
What is an 'AI agent', in the sense that matters for enterprise?+
A system that observes state, chooses an action, executes it against a tool or API, and updates its state - autonomously, over multiple steps, toward a goal. That definition rules out most 'agentic' marketing: a single LLM call with a function schema is a tool-use assistant, not an agent. Agents earn their name when they loop.
Which agent pattern do we actually need?+
Four patterns cover 90% of production wins. Single-tool assistant (one call, one function) for structured extraction. ReAct (reason-act-observe loop) for research and multi-step lookup. Planner-executor (one model plans, another executes) for durable workflows with clear steps. Multi-agent supervisor (a router delegates to specialised agents) for cross-domain work. Escalate one tier at a time - not straight to multi-agent.
What breaks in production that doesn't break in a demo?+
Five things, in order: (1) tool failure - the API returns a 5xx and the agent silently loops. (2) Context blowup - the loop grows the message history until you hit token cost or context limits. (3) Prompt injection - a document the agent reads contains instructions that hijack its plan. (4) Non-determinism - the same input produces different actions across runs. (5) Cost drift - a rare long-tail query costs 200x the median. Each has a specific engineering fix; none are 'solved by a better prompt'.
How do we evaluate an agent before we ship it?+
Trajectory evals, not just answer evals. Score the final answer, but also score whether the agent chose the right tools, in the right order, with the right arguments. Build a fixed test set of 100-500 scenarios with expected trajectories, run it in CI on every prompt or model change, and fail the build on regression. Human review of a stratified sample is non-negotiable for high-stakes paths.
What guardrails are non-optional?+
Six. (1) Max-step limit with a hard timeout. (2) Tool allowlist scoped by user role. (3) Input classification to catch prompt injection before the model sees it. (4) Output schema validation with retry-on-fail. (5) PII redaction at ingress and egress. (6) Per-request cost budget with automatic cutoff. Ship all six on day one; retrofitting them is an order of magnitude harder than building them in.
How much does an enterprise agent cost to run?+
Median ticket cost is the wrong number to track. Track the 95th and 99th percentile - a small fraction of long-running trajectories will dominate the bill. Realistic cost bands, all with GPT-4-class models: structured extraction $0.001-0.01 per call, ReAct research agent $0.05-0.50 per session, planner-executor workflow $0.30-3.00 per run. Below those bands you are probably underinstrumented, not efficient.
Scoping a real agent system?
A 30-minute call with a senior AI architect. We will draw the pattern, the guardrails and the cost curve on the same page.