Harnesses & Agents

What "harness" means — disambiguate first

"Harness" has three overlapping uses in AI interviews. Listen for context:

The agent loop — every framework boils down to this

Internalize this shape. Be able to draw it on a whiteboard.

def agent_loop(user_request, tools, max_steps=20):
    messages = [{"role": "user", "content": user_request}]

    for step in range(max_steps):
        response = llm.complete(messages, tools=tools)

        if response.stop_reason == "end_turn":
            return response.text  # done

        if response.stop_reason == "tool_use":
            tool_results = []
            for tool_call in response.tool_calls:
                result = execute_tool(tool_call.name, tool_call.args)
                tool_results.append({"tool_use_id": tool_call.id, "content": result})

            messages.append({"role": "assistant", "content": response.content})
            messages.append({"role": "user", "content": tool_results})
            continue

    raise MaxStepsExceeded()

Every modern framework — Claude Agent SDK, LangChain agents, OpenAI Assistants, n8n's AI agent node — is a variation of this.

What a production-grade harness adds on top

A naïve while True: call llm() loop is not a harness. A real harness adds:

Agent design patterns — Anthropic's "Building Effective Agents"

Memorize these names. The essay is required reading.

Workflows (deterministic paths)

Agents (open-ended)

Agent — the LLM has tools and decides on its own when to stop. Use only when the task genuinely requires open-endedness; deterministic workflows are cheaper, more predictable, easier to evaluate.

Scaffold vs agent — the tradeoff

For high-stakes compliance, the strong default is scaffold-heavy with surgical agentic steps. You won't deploy a fully open-ended agent that decides whether to file a SAR. You'll deploy a workflow with explicit states, where one step uses an agent to do a bounded subtask.

The major harnesses / frameworks

n8n-specific concepts to know

• Workflow — a graph of nodes executed top-down with branching
• Node — a step (HTTP, transform, AI agent, conditional, wait)
• Trigger node — entry point (cron, webhook, queue)
• Sub-workflows — reusable workflows called from others
• Wait node — pause until human approval, signal, or time
• AI Agent node — LangChain-based LLM with tools
• Memory nodes — chat history, buffer, vector retrieval
• Execution history — every run logged with inputs/outputs at each node — your audit trail substrate

Memory and context management

Real agents accumulate context across steps; production harnesses manage it actively.

• Short-term memory — the conversation/trace itself. The harness decides what to keep in the prompt.
• Long-term memory — vector store + retrieval, explicit memory tools, summarization.
• Working memory / scratchpad — dedicated tool the agent uses to write notes mid-task.
• Compaction / summarization — long traces summarized to fit context windows.
• Sliding window — drop oldest messages when context is full.

10-step sketch — answer any "design an agent" question

Memorize this. You can apply it to any design question.

1. Risk-tier the task. Low (drafts a summary) vs high (recommends a SAR). Different gates.
2. Pick the simplest pattern that fits. Routing → chain → orchestrator-workers → agent.
3. List the tools with descriptions. Mark side-effecting ones. Those need approval.
4. Define the human checkpoints. Where does the workflow pause for sign-off?
5. Specify the audit log. What fields at each step? Where does it persist?
6. Plan the eval. Dataset, grader, what "good" means, offline regression coverage.
7. Plan the failure paths. Tool error, model timeout, ambiguous input, contradictory data.
8. Plan the rollout. Shadow → HITL → autonomous (only for low tier).
9. Plan ongoing monitoring. Online evals, drift, cost/latency dashboards.
10. Plan the kill switch. How to stop the system fast if it's misbehaving.

Talking-point answer: "What's a harness?"

✓ Check yourself

Answer from memory before opening each one.

1. In the canonical agent loop, what two messages do you append after a tool_use stop, and what ends the loop?

   Show answer

   First the assistant message with the model's content blocks, then a user message holding the tool results tagged with their tool_use ids. The loop ends when the model stops with end_turn, or the harness raises once the step limit is reached. See the agent loop.

2. An interviewer asks how you would test your harness. Which of the three meanings is the one that wraps each model call with retries and output repair, and how does it differ from the one that drives the loop?

   Show answer

   That is the LLM harness: retries, parsing, structured-output coercion, caching and logging around a single call. The agent harness is the loop of prompt, tools and results across many calls, and the eval harness scores a system against a dataset. Ask which one they mean before answering. See what harness means.

3. An alert first needs classifying as sanctions, KYC or transaction monitoring, and each class has its own checks. Name the pattern and one reason to prefer it over one open-ended agent.

   Show answer

   Routing: a first call classifies the request and downstream steps differ by class. It is more predictable and cheaper, easier to evaluate per class and easier to audit, because each branch has clear states and its own tools. See workflow patterns.

4. One alert burned forty steps and a large cost because the agent kept repeating the same tool call. Which harness features prevent this, and what should happen at the limit in compliance?

   Show answer

   A hard step or time limit and a token or cost budget stop the runaway, and trace capture lets you see the repetition. In compliance the system halts and escalates to a human; it does not guess. See what a production harness adds.

5. Apply the ten-step sketch to an agent that drafts an impact analysis for a new regulation. Which tier, which pattern and which human checkpoint?

   Show answer

   It is low tier, because the output is an internal draft with no external action. A prompt chain fits: ingest the text, compare it with current controls, draft. The compliance lead reviews before it circulates, the log records the sources and versions, and an eval set from past rule changes measures quality. See the 10-step sketch.

6. A colleague wants an open-ended agent for case investigation because every case is different. What are the trade-offs, and what do you recommend?

   Show answer

   An open-ended agent adapts to novel cases but is costlier, slower, harder to evaluate and harder to audit. Recommend a workflow with explicit states and a bounded agentic step, such as gathering evidence, where steps cannot be listed in advance and a miss is cheap. Keep decisions and filings with people. See scaffold versus agent.

Takeaway

• "Harness" can mean the eval runner, the agent loop or the wrapper around each model call, so ask which before answering.
• Every agent framework is the same loop: call the model, run the tools it asks for, feed the results back, and stop at end of turn or a step limit.
• A production harness adds retries, output repair, budgets, step limits, tool authorization, approval pauses and trace capture.
• Prefer workflows (chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer), and use an open-ended agent only where the steps cannot be listed in advance.
• In high-stakes compliance work, use scaffold with a small agentic step, because it is easier to evaluate and audit.
• Any design answer follows the ten steps, from risk tier through failure paths and rollout to monitoring and a kill switch.

One of the agent's most important tools is retrieval over authoritative documents, and doing that well is its own discipline. → RAG Applied

Further reading

• Building effective agents — Anthropic, 2024. The source of the workflow and agent patterns and the advice to use the simplest one that works.
• Agent SDK overview — Anthropic docs, accessed 2026. The first-party loop with permissions, hooks and sub-agents.
• How we built our multi-agent research system — Anthropic, 2025. An orchestrator-workers design with its failure modes and costs.
• Effective context engineering for AI agents — Anthropic, 2025. How to manage what stays in the prompt as a long trace grows.
• ReAct: Synergizing Reasoning and Acting in Language Models — Yao et al., 2022. The reason-then-act loop behind most agent harnesses.