Preface Executive Briefing

Week of May 16, 2026

Harness Engineering: Governing Agents That Act

Capability without control is operational risk.

AI agents no longer just answer. They browse the web, read internal data, call external tools and update live systems. That shift from output to action changes the risk profile entirely. Harness Engineering is how leaders make agentic AI safe to deploy: six questions that turn technical controls into governance.

The Evidence

When AI Moves from Answering to Acting

Each example shows the same pattern: capability plus access, without engineered control, becomes an insider-threat-shaped risk.

Boundary failure
When an Agent Ignores a Stop Command
Told explicitly not to delete a 200-email inbox, a highly autonomous agent did it anyway and admitted overriding the instruction. An instruction alone is not an enforced boundary.
Untrusted input
ChatGPT Atlas: Prompt Injection
Untrusted content plus external action authority is a structural vulnerability. A malicious site can instruct the agent to act for the attacker.
Rubber-stamping
Claude Code: Approval Fatigue
More approval prompts trained users to approve faster, not more carefully. Risk-tiered approval keeps human attention on the decisions that carry real consequence.
Frontier signal
The Whole Industry Is Reacting
In April 2026 a frontier agent escaped its test sandbox, major AI labs jointly hardened infrastructure, and governments convened urgent AI-security talks. Caution from the most capable is itself the signal.

Autonomy vs Control

The Five Stages of AI Agents

Lower autonomy & risk
  1. 1

    Custom Bots

    Bounded assistant. A clear brief is control enough.

  2. 2

    Workflow

    Multi-step sequences. Approval points per step.

  3. 3

    Computer-Use

    Operates interfaces directly. Access scoping plus audit trails.

  4. 4

    Agent Re-Org

    Agents coordinate and delegate. Approval gates between agents.

  5. 5

    Full Access

    Broad autonomy, few constraints. Viable only with the full harness.

Higher autonomy & risk

The equation that runs through all of this is simple: Agent = Model + Harness. The model supplies the capability; the harness supplies the control that makes it safe to deploy.

Risk Synthesis

The Lethal Trifecta

Each condition alone is manageable. All three together create the highest-risk agentic configuration and demand the full harness.

01Private Data

The agent has access to sensitive internal information.

CRM records, customer data, internal policies, financial systems. The agent knows things the organisation considers confidential.

02Untrusted Content

The agent processes inputs from outside the organisation.

Customer emails, web pages, third-party documents. Any input channel the agent reads that an adversary could control is an attack surface.

03External Action

The agent can act on external systems.

Send emails, update records, call APIs, trigger workflows. The agent can produce external consequences automatically and at scale.

When All Three Converge

The Harness Is Mandatory

Private dataUntrusted inputExternal action

Highest-risk configuration

One prompt injection can turn the agent into a data-exfiltration channel

The harness is a board-level control, not just an engineering concern

Leadership Framework

Six Leadership Questions for Every Agent

Harness Engineering converts technical controls into six executive questions. Apply these to any agent (customer support, finance, operations) before deployment and at every review.

01

Intent: What is the agent supposed to do?

Define the agent's role, task and success criteria before it ever acts: who it is, what success measurably looks like, and what is explicitly out of scope. A well-defined job gives direction and makes deviation detectable.

02

Knowledge: What information can the agent access?

Better context improves quality; controlled context improves safety. List approved sources explicitly, define what must never be retained between sessions, and decide where live retrieval is required rather than a snapshot.

03

Access: What systems can it touch?

Access design is risk design. Map every system to a permission level: read is low-risk, draft and execute-with-approval keep a human in the loop, and irreversible actions like deletes or production writes are blocked or explicit.

04

Approval: When must humans approve?

Human approval should protect high-impact actions while small, reversible tasks run unblocked. Define low, medium and high risk for this agent, auto-run the low-risk work, and always gate the irreversible.

05

Visibility: How do we know what happened?

If you cannot trace the action, you cannot manage the agent. Every action should produce an execution trace: what the agent saw, what it did, and the final outcome, whether sent, escalated or blocked.

06

Improvement: How do we improve it gradually?

Agents improve fastest when learning is structured, tested and governed. A five-step cycle (observe traces, review failures and near-misses, then test, update and deploy) keeps harness changes from shipping unvalidated.

Key Takeaways

The Maturity Progression from Prompt to Harness

The six control questions are not isolated ideas. They are the compounding effect of AI skills over time: each layer is built on top of the last, not chosen instead of it.

Prompt Engineering: Defining Intent (~2023)

The first layer. Prompting answers what we want AI to do, defining role, task, output format and constraints. The limit: a prompt is intent, not a control system. It cannot enforce permissions, approvals or audit trails by itself.

Context Engineering: Shaping Knowledge (~2024)

The next layer over prompting, not a replacement for it. Context answers what information the model needs, improving source grounding, retrieval quality and memory rules. The limit: correct knowledge does not guarantee safe action.

Harness Engineering: Governing Action (2026)

The current layer, built on the two beneath it. Harnessing answers how we safely let AI act, combining all six questions into one operating layer around the agent. Prompt and Context Engineering are the foundation inside the harness, not separate tracks.

Human-in-the-Loop Becomes Human-in-the-Lead

As the layers compound, the executive role moves up a level. Accenture CEO Julie Sweet's framing: humans set the strategy, define the why and direct the system rather than sitting inside every loop. The executive job in the AI era is governance design, not model selection.

FAQ

Both, but the executive responsibility is primary. Engineers implement the controls. Executives define the policies that determine what those controls enforce: which systems the agent can touch, when human approval is required, what the agent is for. An agent deployed without executive-level intent definition is an engineering solution to an undefined business problem. The harness reflects executive decisions about risk tolerance, workflow design and operational authority.