Before an agent changes a budget or sends a message, your team needs a record of the proposed action, enforceable limits and a way to review uncertainty. In my Hard Skill Exchange keynote, I described that structure as a marketing harness.
The keynote covers my approach to coordinating marketing work across channels. This guide explains the controls and turns them into a checklist for your own workflow. The replay is a presentation and Q&A, with slides and illustrated interfaces.
What is an agent harness?
An agent harness is the system around an AI model that supplies context, connects tools, records actions and enforces rules. For marketing, it defines what the agent can prepare or change, what must pause and which evidence a reviewer needs.
In the talk, I organize those controls around a ledger, coded gates and a confidence threshold. The ledger records the touch before it goes out. The gates enforce rules agreed in advance. The threshold routes uncertain decisions to a person. The agentic GTM guide explains the broader marketing approach.
Watch the full keynote
Open the original HSE recording. The official session page identifies the speaker and event. The source recording runs 30:05.
Jump to a section in the original recording
How do you build AI agents for marketing?
Start with one recurring marketing task and draw its path from signal to outcome. Name the data inputs, permitted tools, written limits and review points. Add a record before execution and verify the result at the destination. Expand after reviewing actual outputs.
The keynote describes a sequence spanning intent signals, pages, creative, email, gifting and advertising. Treat each step as a separate action with its own permission and result. A signal can select a workflow; the workflow still needs checks before it sends, spends or publishes.
Turn the keynote’s controls into implementation checks
Scroll the table to compare all columns.
| Control | In the keynote | What to verify in your workflow |
|---|---|---|
| Ledger | Write the proposed touch before execution. | Keep the intended action and the eventual provider result separate. Check which action actually reached its destination. |
| Coded gates | Hold when consent, limits or verification requirements fail. | Test a permitted case and each hold condition. A failed requirement must prevent the action. |
| Human review | Ask a person when the decision falls below the confidence bar. | Give the reviewer the proposed change and supporting evidence. Evaluate the threshold before expanding authority. |
The third column is an implementation checklist, not a claim that these tests ran during the keynote. A tool connection supplies access to an action. Your harness must also govern when that action is allowed. The Metadata MCP guide covers the connection layer.
Keep context current. In the Q&A, I describe using a knowledge vault for active projects, previous work and mistakes. That gives the agent material it can check when it proposes the next action. Sources, dates and unresolved questions should stay visible to the reviewer.
How do you evaluate an AI agent?
Evaluate artifacts, decisions and deliveries separately. Check factual support, targeting, permission checks, reviewer changes and the final destination. Measure business outcomes with a stated attribution rule. A completed run can still produce work that does not help the campaign.
Around 21:00, I discuss a LinkedIn motion that was executing without the desired direct response. Keep that distinction in your reporting. Track the accepted asset, the delivered touch and the commercial result separately. Our pipeline attribution guide explains how to define the final step.
The talk includes reported meeting and optimization totals. It does not provide campaign exports, a controlled comparison or a calibrated model error rate. Use the structure as a design reference, then collect evidence from your own workflow. Our AI marketing workflows guide shows how to assess a draft and its handoff.
Questions from the HSE session
Julia Nimchinski reads audience questions and asks follow-ups. These answers paraphrase the recorded discussion. Navigation times use the original recording clock.
Does the harness need monitoring and alerts?
Yes. In the session, I describe ongoing checks for outbound formatting, language, ad appearance and brand guidelines. These are examples of what the system should inspect. The replay does not demonstrate a live monitoring or alerting implementation.
What is the failure or hallucination rate?
The session does not give a measured failure rate. My answer describes company context, a knowledge vault, model review and vetted skills. Those methods still need evaluation against real decisions. Model agreement and a confidence score do not establish a measured accuracy rate.
How do you approach data enrichment?
I describe a provider waterfall: query an initial source, then use other sources to fill gaps. The results go through a sanity check and email verification. Keep raw contacts separate from contacts ready for a channel. Matching an ad audience does not verify every contact field.
What changes for the CMO?
Around 28:30, I discuss taste, communication and workflow design. Someone still has to choose the audience, decide what deserves attention and define a useful result. Sketch the process first, including its decision points, then use connected tools to build it.
Source and review scope
This guide covers one HSE keynote from the September 8 to 10, 2026 Agentic Harness summit. Gemini 3.8 Flash processed the complete recording’s audio and moving video, including Q&A and closing. Codex verified frame preservation, audio coverage and the source notes. No independent campaign or product test is claimed.