Two B2B teams use the same model with the same prompt. One gets generic output that could belong to any company. The other gets output that names a competitor's launch from last week, reflects the new positioning the leadership team agreed on Monday, and avoids the deprecated product line. The model is the same. The prompt is the same. The variable is what each agent knows about the company it works for.
That gap is what context engineering closes. For marketing agents, it's the difference between a tool that produces work you have to rewrite and one that produces work you can ship.
What context engineering actually means (and the slice this post covers)
Context engineering covers the full set of tokens an LLM sees on each call. System prompts. Tool definitions. Retrieved data. Conversation history. Sub-agent boundaries. Anthropic calls it the natural progression of prompt engineering. If prompt engineering is what you write, context engineering is everything around what you write.
This post is about one slice of that: the business intelligence layer a marketing agent draws on regardless of how it's prompted or what tools it's connected to. Not the system prompt that tells the agent how to behave. Not the tool definitions that tell it what it can do. The relevant facts about the company it's supporting. What the company sells, who buys it, what campaigns are running, what's about to launch. Operating intelligence.
Why marketing agents fail without operating intelligence
Without engineered operating intelligence, marketing agents produce predictable failures:
- Generic copy that could belong to any B2B company.
- Boilerplate optimizations that don’t hit ICPs.
- Data analyses that don’t connect findings to the business.
- Media buying updates based on generic best practices, not your business.
- Content that contradicts a press release the company published last week.
The failure isn't the LLM. The LLM is doing exactly what it's designed to do; it’s generating plausible text against the context it's given. The problem is that the context it's given is whatever a marketer pasted into the prompt that morning, plus whatever it can guess about the company from public web data.
Think about what a new marketing hire learns in their first month: who the company sells to, what's about to launch, what last quarter's positioning shift was, which campaigns are running, who owns which account. Marketing agents start fresh on every run. Without deliberate operating intelligence, they never close that gap.
The four layers of operating intelligence we maintain for every client
When we deploy marketing agents for a client, we maintain four distinct layers of operating intelligence. Each one has its own format and its own audience of agents. Front-loading all four into every agent bloats context and degrades output. We scope each layer to the agents that need it.
Business context
Core context on the company itself. What it does, who it serves, why it exists. Key historical dates like major press releases or funding announcements. Important upcoming events like product releases or conferences.
Format: a markdown file. Clean sections, no fluff.
The piece most teams skip: a guide to the company's operating spine — the core documents the company runs on. A GTM Weekly Leadership Meeting agenda in Notion. A product-launch tracker in Google Drive. A roadmap doc in Linear. We connect those via MCPs and document for the agent when and how to read each one. The agent learns to behave like a teammate who's been in the room.
Every agent we deploy reads the business context. It's the foundation.
Product context
ICP, positioning, messaging pillars, competitive landscape, category language, proof points. The stuff a product marketer would put in a launch doc and then forget to update.
Format: a markdown file with nested tables for the structured pieces like positioning pillars per ICP or a competitor list with one-line differentiators.
Agents that need it: content agents, competitive intel agents, lead-scoring agents. The lead-scoring agent we built for a regulated fintech (covered in Engaging Leads in Hours Instead of Days with a Lead Scoring Agent) reasons over product context to score whether an inbound lead fits the ICP. Is this prospect's company the kind we sell to is a product-context question.
Sales context
Status of active prospects, account owners, deal stages, recent touches. Format: a table or CSV exported from the CRM, plus a guide for the agent to pull live data from CRM MCPs when the table isn't fresh enough. For some clients this is a HubSpot MCP with structured guidance on how to pull sales pipeline. For other clients it's a Google Sheet that gets synced from Salesforce.
Agents that need it: GTM and sales-enablement agents. Outbound writers who need to know which accounts are owned by which rep. Account researchers preparing call prep.
Campaign context
Current and past campaigns: their names in ad accounts, timelines, channels, audiences, objectives, KPIs. Format: a table, plus a guide for pulling live campaign data from ad-platform MCPs.
Agents that need it: anything operating inside a campaign workflow like ad copy generators, creative variant testers, performance reporters. For a B2B startup we work with, an analytics agent uses campaign context to monitor weekly performance across channels. A separate agent uses business, product, and campaign context together to produce a monthly channel-recommendation memo. Different agents, different scopes — same client.
Every agent gets business context. The other three are loaded as needed.
Keeping context fresh: the agent that maintains the agents
Most coverage of context engineering for marketing stops at build the files. The files are the easy part. The hard part is what happens at week three.
A press release lands. A campaign ends. A competitor launches. A positioning pillar shifts. If operating intelligence doesn't keep up, the agents drift, and you're back to rewriting their output.
What works: a purpose-built maintenance agent that runs on a schedule alongside human operators who own each context file (for example, positioning is usually owned by a product-marketing stakeholder). Each run does:
- Internal audit — checks last-update timestamps on connected docs, looks for new campaigns in ad accounts, scans specific Slack channels for announcements.
- External audit — runs targeted web searches on the company, named competitors, and tracked topics.
- Surfaces recommendations — posts specific update suggestions to a single Slack channel ("Competitor X launched feature Y on date Z — update product context table?"). The channel is shared so other team members can see what's being added to operating intelligence.
- Operator approves or edits in Claude.
- Agent writes the approved updates back to the markdown files and tables.
We run business context monthly. The other three layers weekly. Eventually, the maintenance flow can move into a dedicated Slack bot and updates round-tripped directly in Slack.
The maintenance agent is the lever. Every other agent the company runs is downstream of it. A weekly investment in keeping operating intelligence fresh compounds into reliable output across the entire stack.
Where to start
Build the business-context markdown file first. That single file lifts every agent and forces you to write down and organize your thoughts. It’s a good exercise. Then go through the process of connecting it to recurrent workflows involving your AI tools.
Don’t skip the upkeep. Schedule a monthly meeting as a forcing function. Create a prompt you run to audit your business-context markdown file. That’s the first step toward a proper maintenance agent.
Prioritize product, sales, and campaign context based on the needs of your current agents. Go through the same process above.
Always keep operating intelligence in mind when designing AI workflows and agents. It is core to moving from a sprawl of chats to purpose-built agents.
Marketing teams that see agents move KPIs will be the ones who treat operating intelligence as a product, not a doc.
This post draws on anonymized work performed by Lobo Growth across client engagements.