Agent harnesses and inference: a campaign workflow you can review

A campaign can sound excellent and still contain the wrong date, an unsupported promise, or a link that goes nowhere. When you bring AI into marketing, the useful question is: what carries the work from an approved brief to something a person can confidently release?

Two terms help answer that question: inference and agent harness. Understanding them helps you choose a manageable workflow and diagnose where it breaks. You can start with the tools you already use.

Three parts of the system

A model is the trained system that produces outputs from patterns it has learned. Inference is using that model on an input: for example, asking it to turn a campaign brief into an email. Supplying a brand guide as context does not, by itself, retrain the model. Training and fine-tuning are separate processes. Cloudflare's AI glossary explains these distinctions.

An agent harness is the surrounding software that manages an agent's working session. It provides instructions and tools, manages context and state, handles tool requests, and controls how work continues or stops. Microsoft's agent harness guide describes these responsibilities.

An agent might make several inference calls during one campaign task. One call proposes a next step; the harness executes an allowed tool, returns its result, and lets the model continue. A tool failure, missing source, or approval requirement becomes part of the workflow.

Describe: define a reviewable campaign package

Imagine a fictional pottery studio promoting a beginner workshop. The owner wants one email and a landing-page update. This is a planning example, with no claimed client results.

The brief should hold the confirmed date, price, capacity, location, booking URL, audience, and terms. Add two approved examples of the studio's voice and the current page. Identify which document controls when sources disagree. An old social caption should not silently override a newer signed-off brief.

Define the finish line: an email subject line, preview text, email body, revised page copy, and a short check log. Give each asset a filename. Name the reviewer and say what counts as approval. Use the AI Campaign Brief worksheet to collect these inputs before connecting more tools.

Create: give the workflow a small, useful loop

A first version can follow four steps: read the approved materials, draft the assets, compare the draft against the brief, and prepare a review packet. Start with read access to the source folder and permission to write drafts into one destination.

The harness needs enough state to keep track of the campaign ID, brief version, draft locations, unresolved questions, and approval status. That state helps it continue without treating yesterday's draft as today's approved copy. State persistence depends on the product and configuration; confirm how your chosen tool actually handles it.

Make stop conditions concrete. If the date is missing, ask for it. If two sources disagree about the price, mark the conflict. If the tool cannot save a document, report the failure. A polite completion message is not evidence that the file exists.

Starter instruction: Create the five agreed assets from this approved brief. Preserve confirmed facts. Link each factual claim to its source. List missing or conflicting information. Save drafts in the campaign folder, then stop for review. Report the saved files and any checks you could not complete.

The instruction describes the work. Actual tool permissions and approval controls must also be configured in the application. Writing a rule in a prompt does not create a technical access boundary.

Deliver: inspect the assets and the handoff

Review the email and page together. Are their dates and prices identical? Does the booking button point to the approved destination? Does the page make sense on a phone? Does the email preserve any important exclusions? Ask the reviewer to inspect the actual saved assets.

Keep workflow completion separate from campaign performance. A correct, approved package can still receive little interest. Clicks and bookings help evaluate the campaign after release; they do not replace checking what was produced beforehand.

Anthropic's guide to agent evaluations distinguishes a recorded sequence of actions from the outcome in the environment. Apply that principle here: confirm that the files exist and meet the brief, rather than accepting the agent's own statement that everything is ready.

Repeat: test the awkward cases before adding scope

Run three small trials with fictional material: a complete brief, a brief missing a date, and two documents with conflicting prices. Expected results are different. The first should produce a review packet; the others should identify the issue and stop at the relevant boundary.

For each trial, record total time, human review time, corrections, tool failures, and model usage where available. Include repeated calls and retries in the cost. A workflow that drafts quickly but requires extensive repair may offer little practical improvement. Compare it with your current process on the same kind of task.

Start with one agent or a fixed sequence. More coordination becomes useful only when a specific need justifies it. You can add a specialist review step later if you can show what it improves.

Your 20-minute campaign exercise

1. Five minutes: choose one upcoming campaign and write its five required inputs.
2. Five minutes: list the exact assets, destination, owner, and release checkpoint.
3. Five minutes: write three acceptance checks and one condition that should stop the work.
4. Five minutes: run a small draft and inspect the saved result. Record what you corrected.

Bring those notes to the AI Workflow Planner. The next improvement might be a clearer brief, a better tool connection, or a more useful check. Your evidence will tell you where to work.