Problem
Many teams still think of agents as sessions. You start a prompt, the system does something, returns a response and the story ends there.
That model falls short when the work lasts hours or days, needs tools, credentials, persistent context, traces and checkpoints among multiple actors.
The problem is not only intelligence. It’s about the workplace.
Thesis
The acquisition of Ona by OpenAI matters because it names a need that the market was treating as a technical detail: enterprise agents need a persistent workspace.
It is not enough with:
- model
- tools
- context
You also need an environment where the work can continue, save state, touch approved systems and move through reviews without depending on the laptop or the session where it started.
Framework
A serious agentic workspace needs five properties:
- Persistence: the work continues even if the initial session is closed.
- Control: the company decides where it runs and what it can touch.
- Scoped credentials: access with clear and revocable scopes.
- Observability: logs, activity and progress visible.
- Reviewability: handoff to humans without breaking continuity.
Mini-case: a software agent investigates a bug, runs tests, opens a draft fix, waits for feedback and tries again. If everything depends on an ephemeral session, each pause destroys continuity. If it lives in a persistent workspace, it can progress through stages, with checkpoints and real governance.
Measurable signal: percentage of long agentic tasks that can continue without redoing context when the session or device changes.
Posture: the next bottleneck is not thinking better. It’s having a safe place where the agent can keep working.
Why it matters now
OpenAI announced on June 11, 2026 that it plans to acquire Ona to bring its secure execution and cloud orchestration technology to the Codex ecosystem.
The important part of the announcement is not corporate. It’s operational:
- OpenAI says the most valuable work of Codex already happens over hours or days, not minutes.
- It states that companies need agents within environments that meet security, governance and operational requirements.
- It proposes a model where agents run inside the client’s cloud environment, while OpenAI provides intelligence and orchestration.
That is an architectural statement, not just a product one.
Anti-example
“We already have agents because they can execute tasks from a chat window.”
That works for demos or individual work. Not for sustained operations.
If the agent does not have a persistent, controlled and auditable environment, each long task becomes a mix of re‑made context, fuzzy permissions and manual continuity.
Protocol (3 steps)
- Separate session from workspace. A conversation should not be the unit of persistent work.
- Define where the agent lives. Managed cloud, client environment, hybrid or dedicated sandbox.
- Trace continuity and review. The run must survive pauses, owner changes and reviews.
| Layer | Question | Risk if missing |
|---|---|---|
| persistence | the work can continue | rework and drift |
| control | who defines the environment | shadow automation |
| credentials | what can be touched | over‑permissions |
| observability | what was done and when | opacity |
| review | how it is intervened | blockage or blind autonomy |
Related
- Codex on-prem: when software agents leave the public cloud
- Microsoft Foundry Local + Scout: when agent work moves to the perimeter
- Sandboxed Work: the new execution perimeter for production agents
Sources consulted
Next step
Make a simple map of your long agents: which need to keep working when the session ends. If the answer is “almost none”, you are probably still thinking of agents as chat, not as work.
Translated from the Spanish original with AI assistance and reviewed for accuracy. Read the original in Spanish.