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?"