Why Harness, Not Model, Is the Real Key for Deploying AI Agents in Production

OPC Is About a Closed Loop, Not Just a Person

The author defines an OPC (One Person Company) as a tiny unit—either a legal entity or an internal team—where one person makes judgments and a group of agents handles search, preparation, execution, and knowledge capture, all sharing responsibility for the outcome.

The viability of an OPC is judged by whether the end‑to‑end chain runs smoothly, illustrated by the flow diagram:

Business problem
   ↓
Facts & evidence
   ↓
Decision → Preparation → Human confirmation → Execution
                                   ↓
                               Verification / Rollback
                                   ↓
                               Incident / Memory / Skill / Wiki
                                   ↓
                               Reuse next time

Four common failure points in automation are identified:

Context break: New personnel must start from scratch because there is no traceable fact base.

Tool scatter: No single entry point makes troubleshooting chaotic.

Execution invisible: Commands appear successful but lack baseline, verification, or rollback evidence.

Feedback not closed: Repeatedly handling similar issues without writing back results or assets.

To avoid these, the author stresses that the value of a "one‑person ops" setup lies in first closing the loop for a single operator and then handing the reusable path to the team.

Popular Tools Are All Solving the Same Harness Problem

Recent discussions have shifted from "which model is smarter" to "how multiple agents collaborate, connect to tools, and are constrained." The author lists several emerging approaches and their trade‑offs:

Codex Multi‑Agent: Parallel long‑task execution with unified supervision; business priority and acceptance criteria still require human definition.

MCP / A2A: Connects tools, data, and other agents; does not guarantee factual correctness or proper permissions.

Agent Skills: Packages steps and expertise into portable capabilities; quality of judgments inside the skill is not guaranteed.

Kiro / Claude Code Hooks & Permissions: Automatic checks, interceptions, and write‑backs; does not replace human risk decisions.

Anthropic’s analysis of ~400 k Claude Code sessions shows that humans still make most "what to do" decisions, while the model handles "how to do it"; stronger domain expertise lets the model take on more work.

Five Assets That Make an OPC Reusable

The author groups reusable knowledge into five categories:

Authoritative fact sources.

Evidence sufficient to support a judgment.

Steps that can be automated.

Risks that require human confirmation.

Verification, rollback, and write‑back procedures.

These combine into a reusable SOP, workflow, or skill.

Counter‑Intuitive Insights

Do not rush to build a unified RAG. First govern fact sources; use grep and a knowledge base for retrieval, upgrading to RAG only when recall, permission, or freshness consistently fails.

Prompt is not a governance mechanism. Governance must be hardened layer by layer: skills encode steps, hooks intercept and log, permissions restrict actions, and high‑risk steps return to human confirmation.

More knowledge is not always better. An ever‑growing knowledge base can re‑inject stale facts; the author keeps active knowledge in the primary index, periodically reviews, archives by risk, and promotes stable conclusions to the wiki.

Demo: From SOP to a Production‑Ready Skill

A 29‑second recording demonstrates a real production change skill. The skill pre‑defines trigger conditions, a ten‑step process, safety checks, and output format. At runtime the AI prepares commands, diffs, risks, and verification steps, then pauses for human confirmation before execution. After execution it performs smoke tests and writes results back to an incident record and the wiki.

1‑6  AI preparation: context / backup / baseline / diff / pre‑check / risk
7    Human confirmation (no explicit authorization → no execution)
8    Step‑by‑step execution (observable)
9    Result verification (query + smoke test)
10   Write‑back (incident / wiki)

The demo proves that the SOP can be encapsulated as a repeatable skill, not just a one‑off model showcase.

Scaling from Personal OPC to Team Platform

The author visualizes a "π‑shaped" knowledge structure: the left leg covers production/SRE, the right leg covers agent engineering, and the horizontal beam links problem, evidence, execution, verification, and write‑back into a reusable asset.

Progression is described in levels: L4 (personal closed loop) → L5 (team sharing). At L5 the focus shifts from adding more agents to ensuring capabilities are reusable, reviewable, measurable, and publishable.

Three Practical Steps to Start

Select a high‑frequency, low‑risk, verifiable scenario (e.g., alert summary, read‑only inspection, certificate query, CDN self‑service).

Define the five boundaries: source of facts, agent capabilities, moments requiring human involvement, success criteria, and write‑back destination.

Fix a primary execution line first; only after stable repetition add platform features, additional agents, or tool integrations.

The author concludes that a single person can lead an agent army not by wielding more tools, but by providing judgment, discipline, and memory, turning ad‑hoc AI assistance into a reliable production practice.