Agent harnesses and inference: build a client workflow you can review

A client asks for a campaign. The work may include research, an offer, a landing page, promotional copy, a signup journey, and a review. AI can assist across those stages, but the team still needs a process that connects them.

Inference describes the model generating an output from its current input. An agent harness manages the surrounding work loop, including context, tool use, permissions, and progress. Anthropic describes an agent as the model and its harness working together. For a service team, the practical question is how those parts support a deliverable you can stand behind.

Start with a campaign packet that has a clear scope

Consider this fictional assignment: prepare a campaign for a local business workshop. The agreed output is one landing-page draft, three promotional messages, an enquiry-flow review, and a short release checklist. Sending messages and buying ads are separate actions outside this example's drafting scope.

The team supplies approved workshop facts, audience notes, brand guidance, existing page links, and the owner's decisions. A list of missing facts stays visible. The agent should not fill an empty price, date, capacity, or customer quote with an invention.

The campaign packet gives the reviewer something concrete to assess. “Improve our marketing” is a direction; a defined set of assets and checks makes the work manageable.

Map the work before connecting more tools

Stage	Model's contribution	System or team responsibility
Research	Identify useful questions and summarize permitted evidence.	Provide approved sources; preserve source links and dates.
Draft	Generate message options and page copy.	Supply the current brief and maintain a named version.
Build	Propose changes and request editing tools.	Limit access to the agreed project and capture tool results.
Check	Compare the draft with the brief and flag inconsistencies.	Run interaction checks and assign a human reviewer.
Release	Summarize the assets, evidence, and unresolved items.	Approve the exact version and permitted publishing action.

Each new model call uses the information provided at that point. The harness routes tool requests and returns their results. This loop is described in Microsoft's explanation of agent harnesses. A tool being connected does not establish that a task succeeded or that the agent should use it for every assignment.

Make the handoff a deliverable

For the workshop campaign, I would ask the agent to return a small delivery packet containing the page preview, copy variants, a list of confirmed facts, the checks performed, and any unresolved decisions. Every asset should have a clear destination and an owner.

The facts sheet helps a reviewer compare the landing page with the promotional messages. If the page says a place is requested and an ad says a place is guaranteed, the packet contains a consequential inconsistency even when both sentences read well.

The enquiry-flow evidence should identify the environment tested and the result observed. Opening a form is one check. Confirming that an authorized test submission reaches its intended destination is another. Avoid claiming successful delivery from a button click or a generated success message alone.

A campaign delivery checklist to reuse

• Scope: the agreed assets are present, and extra suggestions are clearly separated.
• Facts: the date, audience, price if applicable, location, and commitment match the approved brief.
• Sources: research claims have traceable references, with uncertainty stated.
• Experience: the important page and form interactions have been checked in the relevant environment.
• Review: the reviewer can see the exact version and record accept, revise, or stop.
• Release: publication depends on the configured permission and approval process.
• Handoff: unresolved items, owners, and the next review date are explicit.

Use the list as a starting point and adapt it to the assignment. A draft-only task needs different access and release controls from an agent authorized to publish a page.

Try three exceptions before calling the workflow repeatable

Missing input: remove the workshop date from a practice brief. The expected behavior is a visible question or an explicitly incomplete draft. A fabricated date should fail review.

Tool failure: use a controlled test where a preview or test destination is unavailable. The agent should report which step failed, preserve completed work, and avoid claiming the whole campaign is ready.

Changed approval: revise an important claim after a draft was approved. The release process should require review of the changed version. An approval for an earlier version should not silently authorize every later edit.

These are proposed checks for a workflow you configure. They are not claims that every agent product supports the same controls. Look for actual permission settings, execution limits, saved state, and review behavior in the environment you select.

Measure the complete job

Record preparation, production, review, corrections, and any recovery after a failed step. Record model and tool usage when the platform exposes it. Then ask whether the final packet met the same quality standard as your usual process.

Useful measures include the share of packets accepted without substantial correction, unsupported claims found during review, missing assets, elapsed delivery time, and total team time. Keep the sample small enough to inspect. Early runs can guide your next decision without proving a universal productivity claim.

Inference is ordinary model use, rather than a new training process for each assignment; Cloudflare's explanation provides the technical background. More model calls create more work to process and review. Set a reasonable scope and stop condition before adding extra rounds or specialist agents.

A practical brief for your first run

Prepare [campaign packet] for [audience] using [approved facts and sources]. Produce [specific assets]. Preserve [brand and business decisions]. List missing information before relying on it. Use [permitted tools] within [scope and limits]. Return the assets, their destinations, the checks performed, and unresolved items. Keep [release actions] dependent on approval of the exact version.

Save the brief with the actual tools, permissions, and review instructions. A saved prompt alone does not schedule future work or create an enforceable approval step.

Give one useful workflow a clear owner

Choose one recurring deliverable and run it on a permitted practice example. Review the packet with the person who normally signs off on that work. Improve the missing input, tool connection, or check that created the most rework before expanding the process.

Read how we organize AI-assisted client work for the broader delivery approach, or apply a similar review process to client reporting. To practice the foundations, explore AI Teachers' first useful workflow course.

Describe. Create. Deliver. Repeat. Make each repetition easier to understand, review, and improve.

Published November 5, 2026. The campaign, checklist, and exception drills are original illustrative exercises, not measured client results. Technical references were reviewed for this article.