A complete guide to getting started with agentic content operations
Feeding content agents the right context, real-world use cases, and what to decide before you get started
Kelsey Bernius
Staff Enterprise Product Marketing Manager at Sanity
Published
AI projects often start off full of ambition, alignment, and positive energy. But when teams are debating models, workflows, and processes, the foundational content that feeds the agents is often overlooked. And agents can spit out some wily outputs as a result.
The city of New York City learned this lesson all too well when they rolled out an AI-chatbot in 2023 that was supposed to provide new and existing business owners with reliable information and answers about running a business in the city. Instead the chatbot told business owners they could take a cut of worker tips, fire people who complained about harassment, and refuse tenants with housing vouchers. Whoops. While the source material was official, and in some cases partially trure, the content was spread across many pages and agencies with no structure linking each rule to its context. And the bot often stitched it together wrong.
Poor outputs run completely counter to the whole premise of AI: save time and resources. “The number one barrier for good performance, it’s not cost, it’s quality of output,” says Sanity CEO Magnus Hillestad. “And off the shelf agents often fail because they don’t understand your content, your context, so they fail to tell your story.”
This guide explains how to provide proper context and reliable content for your agents, shares real-world agentic
Read the full guide
While discovery is how an agent reaches the content, content is the material itself, and context is the additional instructions, framework, timing considerations, or environments. Context and content are so tightly bound today that they can be considered the same within content operations.
Most companies store both material content and the supporting context across a marketing website, Google Drive, maybe an abandoned Confluence page, and esoteric knowledge found in the heads of whoever happened to be in the room when certain decisions were made.
Over time, context gets scattered across docs, tools and teams, and it becomes hard for anyone to navigate. An agent can reason over whatever you hand it, but fragmented context means it starts every task like a new research project, rebuilding what it needs from scratch. That burns tokens like there’s no tomorrow, and it also leaves more room for the agent to miss something or fill gaps with guesses
Before you send content and context to your agents, consider the following readiness checklist:
- Narrow the scope: Name the specific task the agent will do, the content types it will touch, and the systems it needs. An agent drafting product pages needs different context than one tagging or routing support tickets. Loading everything might feel like the right thing to do, but irrelevant documents water down retrieval and make wrong answers more likely.
- Structure the content before you hand it over: How you structure your content will be tailored to your business. A high level example of
structured content - Designate when agents cannot guess: List out content such as pricing, product names, release dates, legal and compliance claims that agents can never assume or make guesses on. These come from a source of record every time, never from the model's training data.
- Describe fields in plain language: The help text that tells an editor what a field is for also tells an agent what it means and when to use it. Content models without descriptions leave the agent to infer intent.
- Package rules as reusable instructions: Brand voice, style guide, banned claims and approval requirements belong in one maintained place that every agent reads. If each person pastes them into their own prompts, you end up with five slightly different versions of the brand.
- Scope permissions to match the task: Give read access to what the agent needs and write access only after the output has earned trust from a human review step.
- Keep context connected to the source: Stale context produces confident, but wrong answers. When agents pull from the same content your site publishes from, one update reaches both.
- Test with real tasks and trace every miss. When an output fails, find the missing or incorrect context behind it and fix the source rather than the prompt.
While a good checklist helps you frame your approach, seeing how this plays out in real companies brings agentic content operations from a theory you read about to something you implement. The following Sanity customers are some of our favorite use cases of agentic content operations.
Sigma's marketing team is five people, and they support a go-to-market organization of more than 200. With resources sparse, the marketing team wanted to start using agents to post content for the team. Luke Stanke, who leads marketing at Sigma, knew that the fix had to start with structured content before introducing agents.
"We can go fast being able to build out everything, but we need to have something that is structured behind the scenes where we could go pull from at all times," said Stanke.
The Sigma site now runs on 128 section types across 11 page categories, and they all draw on one shared page builder. Instead of 11 different page building systems, there is one shared vocabulary for building pages.
Every section includes a:
- Schema, which defines the valid fields and the shape of the data. It lives in Sanity, so content can be validated against it and the agent can check its own output.
- Readme, which as a plain text file explains what the section is for, when to use it, what props it takes, and what good usage looks like.
- Screenshot, which shows what the section looks like. It serves the agent as a visual reference and the editor as a preview in the page builder, so nobody is choosing a section based on its name.
- Skill, which covers how to use or build a certain page section.
The page builder agent reads the prompt, works out what kind of page it is, and composes it from sections that already exist, checking each block against the schema, the readme, and the screenshot. When no existing section fits, the “create section” skill builds a new one. It doesn’t improvise a one-off for that page. It creates the full set: component, schema, readme, screenshot, and CI checks and tests. The new section then joins the shared page builder, which means it is available on all 11 page types from that moment on.
Stanke and his team now builds an ABM landing page from a single prompt. It reads Sigma's own telemetry on similar customers, pulls the Wistia video transcripts, matches them to the topic, and places the right videos itself. The pages are anonymized and kept out of search.
"It's not going to be perfect by default," says Luke. A page takes about five minutes to deploy, and the cleanup after it runs under two minutes, mostly checking for the obvious blunder like a wrong company name. Normally the page goes to staging and someone reviews it before it ships.
The team built a second
Takeaway: An agent can only build what it can understand. Sigma put the structure in Sanity so the agent always knows what a valid page is made of, then documented what each of their 128 sections is for so it picks the right one. The result isn’t a website that’s faster to edit. It’s one that 40 sales reps and agents can build on, from the same section library and the same structured content.
Digital communications company Braze first brought in Sanity for a customer education and marketing website. But as agents became more capable, the team started applying them across teams and applications.
Matt Chamberlain, who leads applied AI product at Braze, took that idea to internal content. The highest-stakes internal content he could find was competitive intelligence, the guidance that tells a sales rep how to win a specific deal against a specific competitor. This type of content is carefully written, kept current by people who know the market, and as he says, almost impossible to find when you need it.
"There's all of this brilliant, lovingly maintained content that no one knows where it exists," he told the room. "I hope you're not sick that day, because otherwise you're never going to know how you're going to win against Salesforce."
Braze built the tool, called Flux Ops, in about a day, and includes the following components:
- The
front end - The team uses Sanity Knowledge Bases (now in beta) to turn its competitive intel into short, plain-text fact sheets that any of its agents can look up and read. The Knowledge Bases provide context specifically built for agents from a defined purpose. The ingest process also reads across sources and surfaces contradictions and factual gaps for a person to resolve, which means the system tells you where your own documentation disagrees with itself.
- An agent ingests a Gong sales call transcript, extracts signal (claims made about Braze, claims made about competitors, objections raised, concerns voiced), checks the rep's handling against the documented competitive playbook, validates product claims against public documentation and competitors' own sites, and produces findings. Some findings are things the rep did well. Some are coachable moments. Some are gaps in the intel itself, where a prospect raised an objection nobody had documented a response to.
- Sanity Functions provide the runtime, triggered by changes to the workflow document.
A human reviews each output before anything reaches the core dataset. A reviewer can approve it, correct it, or reject it, and the interface links directly to the moment in the transcript where the agent thinks the thing happened, so a claim can be easily checked.
"It's critical that it's accurate and that it's honed with care, and that we're not just leaving it up to agents to figure out how to communicate on behalf of our brand,” said Chamberlain. “That's never going to fly."
Human reviewers can fix a single output, or they can write a preference that gets fed back into every future review. The same pattern shows up in Knowledge Bases, where resolving a contradiction stores an instruction that applies on every rebuild. Review effort compounds instead of being spent once.
Watch Matt’s entire session at Everything *[NYC].
Takeaway: Content operations is not only for the content team. Almost every team in a company runs a content operation of some kind, whether or not they call it that. Start where the content is valuable, but hard to find. Agents are capable if you point them in the correct location.
MJH Life Sciences runs 60 medical publishing brands on Sanity, serving oncologists, neurologists, and dermatologists. Their content goes further than a website. MJH pushes it into electronic health record systems, so a doctor treating a patient can see MJH material on tumors, genetic markers, and biomarkers in the side panel of the record, with Sanity behind it.
Andy Martin, MJH’s CTO, knows precisely who he is publishing for. Describing CGTLive, one of the sixty properties, he did the arithmetic himself: "I came up with 12,525 individual people who would care about this website. So we're out here making a website for those 12,525 people."
At that level of specificity, traditional content metrics stop being useful. Traffic only tells you a small piece about what’s going well for your content. What Andy wanted to know was whether the content was landing with the people it was written for.
Andy and his team used an experimental Sanity feature called Resonance, that reads a site the way its audiences would. Resonance builds a model of the business and its audiences, simulates those audiences against the content, and surfaces the gaps.
Resonance found that the MJH site was not serving clinical investigators well. It took the perspective of a biotech strategy professional, rather than of an investigator leading a trial, and brought the voice of the customer onto the site rather than leaving it in a persona document.
The team made two major changes to its content:
Therapy status boxes now sit at the top of articles. Traffic logs showed readers arriving on a single news item out of sequence, with no way to tell whether they were looking at current news or something from two years ago. The box gives the arc of the drug: when it was discovered, which firm is developing it, which trial phases it has passed, where it stands today.
A development watch replaced the usual trending-articles module in the navigation, showing where each drug sits in the stages of clinical development. A reader arriving from search now lands somewhere that orients them.
"We never could have thought about it, we didn't think about it, we weren't doing it, and our audience and our site is now better for it," said Andy.
The examples above cover basic publishing processes to real-time
Who approves what?
Braze's model is a good starting point. Agents draft. Humans approve anything that reaches the core dataset. The review interface links to the exact moment in the source so a reviewer can check the claim rather than take it on faith. Decide where your stopping points are before an agent has write access to anything that matters, and write the rule down somewhere people will find it.
What makes the experience trustworthy?
Ability is whether the thing can actually do the work. Transparency is whether a reviewer can see where a suggestion came from. Control is whether a person can accept, decline, and correct. Most agent projects ship the first and skip the others, which is why people stop trusting them three month in.
Where does clarity live?
Even Westvang, Sanity co-founder, argues that writing clear instructions for agents is not a software engineering responsibility. Marketing owns the tone of voice. Product owns how products are described. Legal reads the agent's
What are you willing to expose?
An agent that reads all of your content will find every gap and contradiction in it. Decide whether you would rather find those gaps yourself, on your own schedule, or have a customer find them for you.
Where does the work run?
Chamberlain’svpoint about vendor choice applies to more than agents. His team knew what they wanted to build. What determined whether it happened in a day or a quarter was whether one set of APIs and an opinionated framework let them act on an idea they already knew was good, rather than convening three teams to agree on a runtime first.
“Agents don’t do simplicity,” points out Sanity CTO Simen Svale Skogsrud. “That’s still our job.”
Find the content in your organization that people ask for by name and cannot locate. That is where the valuable building can and should start.
Ready to talk to an expert about AI content operations? Let's do it.