Companies are putting AI agents in front of customers everywhere: support, sales, in-product. Each of these agents usually start a conversation knowing two things. It knows who the customer is, from your CRM. It knows what your business offers, from a knowledge base. But it doesn't know the one thing that matters most in the moment: what the customer is actually doing. 

Every customer has a journey. They've been comparing options, retracing steps, hesitating over a decision. The moment they reach the agent, that history is invisible, so the agent responds to the message it was sent rather than the journey behind it. It guesses.

Today we're releasing Agentic Context, a real-time capability in Snowplow Signals that closes that gap. It gives your agent a ready-to-use view of what a customer is doing right now: a pre-processed, ordered summary of their recent behavioral events, filtered, sequenced, and formatted to drop straight into the agent's context window.

Real-time context has been our argument for what separates a useful agent from a generic one. Agentic Context delivers it: a live view of customer behavior your agent can consume the moment it is needed.

An attribute is a fact. The event stream is intent.

Most tools that claim real-time context serve a single attribute value: a score, a flag, a count. Engagement: high. Sessions this week: four. That's a snapshot, and it's useful for some things. But it can't tell an agent that the customer compared two products, went back to one of them twice, paused on the details, and has been hesitating for over a minute.

An attribute tells the agent what is true. The event stream shows what the customer is trying to do, as it unfolds. You can't reconstruct a sequence, or the timing between the steps, from a single rolled-up value. And the timing is signal in its own right: rushing or hesitating, decisive or stuck.

“A semantic layer describes the past consistently; a context layer assembles the present for a decision.”

Yali Sassoon, Snowplow CTO & Co-Founder,  A semantic layer is not a context layer

Why other systems can't do this

Customer data platforms are built around a person-centric, relational model. They store attributes about a user and roll them up over time. That design can answer “what is this customer's lifetime value,” but not “what has this customer done in the last ninety seconds, and in what order.” The event-by-event sequence isn't something they keep in a form an agent can read in the moment.

That gap has a cost in the conversation. The agent asks the customer to repeat themselves, misses the moment to intervene, or hands off a problem it had enough information to solve. Every one of those is a point where conversion, resolution time, or retention takes the hit. It's already showing with today's context-blind agents, with 75% of consumers saying AI customer service leaves them frustrated (Glance).

Snowplow is built on the behavioral event stream from the ground up. Every action is captured as a structured, timed event, which is why we can hand an agent the sequence itself instead of a summary of it. No other vendor delivers the ordered, in-session behavioral stream shaped for an agent's context window. This is the differentiator we've talked about for years, now made concrete in a feature you can put in front of an agent.

What your agent can do once it can see

Lift conversion by acting on hesitation. When the agent can see a customer stalling at a decision point, bouncing between options, or returning to the same item, it can step in at the moment it matters and help move them from “almost” to “bought.”

Resolve issues without re-explaining. The agent already knows what the customer was doing and where it went wrong, so it picks up mid-problem and goes straight to the fix, instead of asking them to walk through what happened. 

Guide customers who are stuck. When someone loops on the same step or can't find what they need, the agent recognizes the pattern and offers help that fits the moment, before frustration turns into a drop-off.

We've posted a short demo where an agent watches a live e-commerce session, notices the customer hesitating in the cart just short of the free-shipping threshold, and offers relevant help. There's no hand-written rule for that scenario; the agent is reading the event log and reasoning on it. Watch the demo →

How it works

Your agent fetches context with a single API call using a session ID, either the moment it's invoked or mid-conversation as a tool call. You choose which events and level of detail the agent should see, and Snowplow handles the filtering, ordering, timing, and formatting. There's no transformation logic to write and no serving layer to stand up, so your team ships the agent faster and doesn’t take on another data pipeline to maintain. Full setup, in the Console or the Python SDK, is in the Agentic Context docs.

Agentic Context doesn’t replace attribute groups; it works alongside them. Historical profile data and real-time in-session behavior can be retrieved together, so the agent gets both who the customer is and what they're doing right now in a single view.

It's vendor-neutral. Fetch context over the API or as an MCP tool call, from whatever agent framework you're building on.

Available today

Agentic Context is live now and available in our free trial, so you can try it against your own data. It requires Signals, and works with the behavioral data your Snowplow SDKs already capture across web, mobile, and server-side.