The distinction that keeps getting blurred
"Agentic" gets applied to two very different things. An AI agent is the actor: a model that plans and takes actions — Claude, ChatGPT, a custom harness. Agentic software is the acted-upon: a product whose capabilities are exposed so agents can do real work through it. The distinction matters because they're built by different people. Frontier agents come from model labs with capital measured in billions. Agentic software is what everyone else can — and should — build.
Our position, stated plainly: a martech company is not going to out-build Anthropic, OpenAI, or Google on models. The durable work is the tooling around the model — exposing your product through MCP, a CLI, and APIs so a capable AI can drive your system and your customers can compose it into their own workflows.
The test for whether software is actually agentic
Ask what an agent can do through the product's public tool surface, end to end, without a human clicking approve at every step. Can it read real state? Take real actions? Verify the result? If the answer is "there's an AI assistant in our UI," that's a chat box, not agentic software — the agent is trapped inside the vendor's walls.
The inverse failure is agent washing: shipping thirteen named "agents" that are thin prompt wrappers around someone else's model, bolted onto a product that needed an AI story. The agent count is marketing. What the agents are wired into — the tool surface — tells you everything.
What it looks like in practice
AXO is built as agentic software. Every operation in the personalization layer — reading segments and analytics, drafting and staging variants, wiring placements, firing triggers, managing profiles, tuning the bandit — is one of 130+ MCP tools, also reachable over REST. An agent connects with a scoped token, does the work, and the human decides where the agent runs autonomously and where it asks first.