Agent harnesses and inference: deliver a small AI service consistently
Understand the system behind an AI-assisted service. Scope a content package, add review checkpoints, and calculate the time and cost of delivering it.
Your first AI service needs a clear finish line. A buyer should understand what they provide, what they receive, and how you will check it. Understanding inference and an agent harness can help you make that delivery process easier to repeat.
Consider a fictional service: turn one approved 20-minute recording into three social posts and one email, with one consolidated revision. You are selling a reviewed content package. The workflow behind it needs to support that promise.
What happens when you ask AI to do the work?
A model is a trained system. Inference is the process of applying it to an input to produce an output. Each time you ask for a draft or a revision, you are using inference. Giving the model a transcript and brand examples supplies context; it does not, by itself, train the model. Cloudflare's glossary distinguishes inference from training and fine-tuning.
An agent harness is the software around the model that manages the work: instructions, tools, context, state, and the loop between model requests and tool results. Microsoft's explanation describes how those pieces support a working agent.
For your service, the harness might load an approved transcript, ask the model for an outline, save draft files, and return a checklist for review. Several inference calls can belong to one job. The harness does not make every output correct; it gives you places to structure, inspect, and control the process.
Describe the offer before automating it
Write the scope as if another person will deliver it tomorrow. Include the recording length limit, required audio quality, audience, intended channels, length of each asset, revision allowance, and delivery format. Confirm the client's permission to use the source material in the tools you plan to use.
Ask for the current offer, approved links, and two voice examples. If the recording mentions an expired offer, the workflow should flag it. If a speaker uses a customer's name, confirm whether it belongs in public copy. These are delivery decisions you need to resolve.
A sample definition of done: three edited posts, one edited email, source references for factual claims, and a review log delivered in an agreed folder. Each asset is labeled “Ready for client review.” Publishing happens only if the agreed scope and approval cover it.
Use the First Offer Builder to turn that scope into a short offer. Keep the promise within work you can personally evaluate.
Create a workflow you can inspect
Begin with a small sequence. Load the source; extract candidate ideas; draft the four assets; check them against the recording; prepare the handoff. You can perform the early version manually using an AI assistant and a checklist, then automate the parts that repeat reliably.
When tools are connected, keep access appropriate to this job. Reading the approved source folder and creating draft documents may be enough. Set actual permissions in the application. A prompt saying “draft only” is useful instruction, but it does not substitute for an access control.
Retain a simple job record: source version, selected excerpts, draft filenames, unanswered questions, client feedback, and current status. Useful statuses might be “Input needed,” “Drafting,” “Review needed,” and “Delivered.” Choose a system that preserves the information you need between sessions.
For each asset, ask the model to identify supporting transcript passages or timestamps. Then check those references yourself. If a quote is unclear, leave it out or ask the client. Do not let the model invent a smoother quotation.
Deliver evidence of the work
A completion message is easy to generate. A usable package needs actual files, correct links, accurate claims, and a clear handoff. Open each file before delivery. Check that the four assets are distinct, fit their intended channels, and reflect what the speaker actually said.
Anthropic's evaluation guide emphasizes checking the outcome in the environment as well as the agent's recorded actions. For this service, a review log should say which checks passed, which corrections you made, and which questions remain.
A client should be able to tell what is ready and what still needs a decision. If a document failed to save, fix it before marking the package delivered. If a client approves changes, preserve the approved version so the next revision does not reintroduce an old claim.
Count the full delivery cost
Inference is one part of the cost. Preparation, transcript cleanup, editing, client communication, revisions, software, and repeated model calls also matter. Measure a sample before deciding what price or turnaround you can sustain.
Here is an illustrative planning calculation, using invented assumptions rather than market rates:
| Work | Example time |
|---|---|
| Brief and source preparation | 25 minutes |
| AI drafting and tool handling | 20 minutes |
| Factual and editorial review | 45 minutes |
| Revision and client communication | 30 minutes |
| Final files and handoff | 15 minutes |
That totals 135 minutes, or 2.25 hours. At an illustrative internal labor allowance of $40 per hour, plus $6 of allocated software and usage costs, the delivery estimate is $96. It is not a recommended price or profit forecast. It excludes acquisition time, wider overhead, taxes, and contingency.
If the job actually takes three hours, the same calculation becomes $126. That difference is a reason to record review time and clarify the scope. Use your own assumptions in the first-service pricing guide.
Repeat with three sample jobs
Try a clear recording, a recording with an outdated offer, and a recording that lacks enough useful material. The workflow should not produce the same confident response to all three. Define when it should proceed, request clarification, or recommend a smaller deliverable.
Compare corrections, elapsed time, hands-on time, and total usage. A more elaborate agent setup earns its place only if it improves a result you care about. One well-scoped assistant can be enough for a first service.
Your 30-minute next step
- Choose a recording you own or have permission to reuse.
- Spend ten minutes writing the scope, required inputs, and checks.
- Spend fifteen minutes making one sample asset and tracing its claims back to the source.
- Spend five minutes logging corrections and estimating the remaining work.
Label the sample as a demonstration. Use it to ask a suitable buyer whether the deliverable solves a real problem, without claiming client results you have not earned. The content repurposing guide helps you develop the full package once the scope is clear.