Agents Need a Runtime
Companies are not short of agent demos. They are short of agent operating systems. A useful enterprise agent needs permission boundaries, tool catalogs, memory policy, logs, evals, human escalation, data access rules, and a way to explain what it did after the fact.
That is why managed-agent announcements matter. They signal that the market is moving from raw model access toward governed execution. The hard problem is no longer asking a model to draft an email. It is allowing an agent to touch production systems without becoming a compliance incident.
The Missing Layers Are Boring
The missing layers are not glamorous. Identity, audit trails, sandboxing, test environments, cost caps, rollback, and model-change monitoring rarely appear in keynote demos. They decide whether the pilot survives procurement.
| Reader question | What matters now | Editorial answer |
|---|---|---|
| What fails first? | Unclear permissions | Agents need explicit authority. |
| What proves value? | Workflow evals | Test actual work, not generic prompts. |
| What keeps trust? | Auditability | Every action needs a trail. |
Buy the Harness, Not the Demo
Before buying agents, companies should map the workflow, classify every tool by risk, define which actions need approval, and build evals around real failure modes. The agent should enter a system that already knows what safe work means.
Do not buy an agent until you know which actions it may take, which actions it may suggest, and which actions it must never touch.
The better enterprise question is not which agent is smartest. It is which agent can be trusted inside the company's existing controls.