Three Vendors, One Week, One Verdict: The Chatbot Needs a Production Layer, Not a Bigger Model

Three separate companies put out three separate announcements this week, and none of them mention each other. OpenAI shipped an enterprise platform for agents that finish tasks instead of describing them. Alibaba Cloud unveiled an agent-lifecycle platform built for retries and audits rather than clever prompts. And Akamai published a security report arguing that the access-control model built for human employees doesn't work for the agents now acting on their behalf. Read separately, they're a product launch, an infrastructure roadmap, and a security brief. Read together, they're the same admission from three directions: once a chatbot starts taking actions instead of just answering, the model stops being the interesting part, and the scaffolding around it becomes the whole job.

1. OpenAI Presence: Completing the Transaction, Not Explaining It

OpenAI introduced Presence this month as a managed platform for deploying real-time voice and chat agents that resolve requests directly — checking an account, rebooking a flight, issuing a refund — rather than describing the policy and pointing the user somewhere else to finish it. The distinction OpenAI is drawing matters more than the feature list: a conventional support chatbot's job ends at a correct answer, while Presence's job ends at a completed transaction, which means the model's output has to be paired with policies, guardrails, and escalation rules that a plain chat interface never needed, plus monitoring to catch it when those rules aren't enough.

That's a familiar shift for any team that has moved a bot from "answer questions" to "call tools." The moment an agent can actually change account state, correctness stops being the only bar — reversibility, auditability, and a clean human handoff become requirements, not nice-to-haves. Presence packaging those as a managed product, rather than leaving each enterprise to bolt them on, is itself a signal of how much of this work has turned out to be repeatable plumbing rather than a business-specific problem.

2. Alibaba's AgentCore: The Lifecycle Is the Product

At its Hangzhou Cloud Summit, Alibaba Cloud unveiled AgentCore, an enterprise platform for building, running, and managing agents across their full lifecycle — not a model, but the infrastructure underneath one: long-running task support, automatic retry, checkpoint recovery, asynchronous execution, isolated execution environments, and end-to-end audit trails for the models, MCP servers, and skills an agent draws on. Notably, none of that list is about reasoning quality. It's the same category of problem a backend team solves for any long-running job queue, applied to agents because agents turned out to need it too — an agent that fails midway through a multi-step task and can't checkpoint or retry cleanly is a reliability bug, whatever the model underneath it can do.

// The gap AgentCore-style infrastructure targets: what a bare tool-calling
// loop is missing once a task spans multiple steps and can fail partway.
async function runAgentTask(task: Task) {
  const checkpoint = await loadCheckpoint(task.id); // resume, don't restart
  for (const step of task.remainingSteps(checkpoint)) {
    try {
      await executeStep(step);
      await saveCheckpoint(task.id, step);          // survive a crash mid-task
    } catch (err) {
      if (isRetryable(err)) continue;                // bounded automatic retry
      await escalateToHuman(task, step, err);         // don't fail silently
      return;
    }
  }
}

3. Akamai's Warning: Access Control Was Built for People, Not Agents

Akamai's latest State of the Internet security report, "Speed, Scale, and Nonhuman Identity," makes the governance case for why platforms like the two above matter. Among its findings: more than 6% of chatbot conversations sampled contained sensitive information, 47% of enterprise chatbot interactions ran through unmonitored personal accounts rather than managed enterprise ones, and over 40% of enterprise users had installed AI browser extensions, a quarter of which changed their own permissions within a year of install. Akamai's argument is that identity and access management, built around verifying who a human is, doesn't transfer cleanly to governing what an agent does — the report calls for matching an agent's autonomy to how verifiable and reversible its actions are, keeping a human in the loop wherever a mistake is hard to undo.

That framing is the missing piece connecting the other two announcements: infrastructure like AgentCore and products like Presence give an agent the ability to act reliably, but reliability isn't the same as safety. A checkpointed, auditable agent can still be handed too much autonomy for an action nobody can easily reverse. The report's autonomy-to-verifiability principle is a reasonable default for any team wiring tool calls into a chat agent, regardless of whose infrastructure it runs on.

What Connects Them

None of these three companies is claiming a model breakthrough this week, and that's the point. OpenAI is selling the runtime around the model, Alibaba is selling the lifecycle infrastructure underneath it, and Akamai is measuring what goes wrong when neither exists. A chatbot that only answers questions never needed retry logic, audit trails, or a policy about which actions require a human — a chatbot that acts does, immediately, and it turns out three very different vendors converged on that same list within the same week.

Suggested visuals for this post: a three-column comparison graphic mapping OpenAI Presence, Alibaba AgentCore, and Akamai's report onto "execution," "lifecycle," and "governance" layers of an agent stack; a flowchart of the autonomy-to-verifiability decision Akamai recommends for gating a tool call; and an annotated version of the checkpoint/retry code snippet showing where each vendor's product would plug in.

— Maya