Prototype · closed beta
unOpen.AI
In closed beta: Roxy, a private AI assistant with permanent memory running on infrastructure the user owns. Brand, site, product, backend and conversational design.
Technology · September 2026
Agents that carry
the work forward.
An AI agent runtime runs the loop around a model: continuing a multi-step task, maintaining context, recovering from failures and coordinating tools. Managed runtimes such as OpenAI’s Agents API, in public beta since September 2026, take that loop off your engineering team. They suit long-running enterprise workflows; you still define success, the data and which actions need review.
Scope an agent pilot
All Technology & Model Stack →
OpenAI Agents API · Public beta · Reviewed 19 September 2026
Introduction transcript. Qubi: Some jobs take many steps, and a model on its own loses its place. A managed runtime keeps the session going: it remembers, tidies its notes, and recovers. Tools run inside a sandbox, and any change to outside systems waits for approval. Your team reviews a checked result, then accepts it or lets it continue.
What changed
The execution loop becomes a service.
OpenAI released the Agents API in public beta on 10 September 2026. It manages durable sessions, orchestration, context compaction and recovery, with hosted or connected sandbox environments. Official release changelog.
The runtime can execute code, work with files, connect to Model Context Protocol (MCP) servers and return artefacts. A session can continue across turns, so an application can show progress and let a user steer the work. Agents API documentation.
The business value to test is reduced engineering effort around a task that genuinely needs several steps. Your team still defines what success means, which data is available and which actions require review. A managed runtime does not make an incomplete brief or unreliable source data disappear.
Illustrative workflow
AN INVESTIGATION WITH A REVIEWABLE RESULT
-
01Request
scope + success criteria
-
02Session
bounded working context
-
03Tools
approved data + sandbox
-
04Verification
tests + source checks
-
05Review
accept or continue
Where to start
01
An incident investigator
Start with a redacted alert, read-only logs and a runbook. Produce a timeline, likely causes and evidence links. A human owns the decision to change a live service.
02
A document reviewer
Compare a supplier submission with an approved requirements checklist. Return missing evidence and page references. Keep acceptance criteria outside the model and ask a reviewer to resolve disputed findings.
03
A business-data analyst
Answer a bounded question against approved warehouse views. Return the query, metric definitions and a chart. Reconcile totals to a trusted report before sharing the explanation.
Worked example
Why did support demand rise?
Proposed pilot: an operations manager supplies a date range and asks why ticket volume increased. The agent reads permitted ticket summaries and product-release notes, groups the changes, and produces a source-linked explanation. Customer identifiers are removed unless needed for the approved question.
Verification runs outside the narrative: ticket counts must match the reporting system, categories must sum correctly, and every claimed release-related cause needs supporting evidence. Correlation is presented as a hypothesis until checked. Missing data produces an explicit limitation and a request for the missing source.
Evaluate time to a reviewed answer, arithmetic errors, unsupported explanations and total cost. Compare with a straightforward saved query and dashboard. The agent earns its place when the question varies enough that the fixed workflow becomes burdensome.
Deployment and economics
Choose who runs each part.
OpenAI’s documentation separates the managed session from the execution environment. A self-hosted sandbox is not a self-hosted model or control plane. Map which prompts, tool results and artefacts cross each boundary before selecting the deployment. Runtime and environment overview.
Model tokens, tools and hosted containers are separate cost lines. Price a complete job, including waiting time, retries and review. Set time limits, stop abandoned environments and record spend by task. Published billing categories. Our AI observability guide explains the wider operational measurement.
Use a conventional worker for deterministic jobs. For cross-platform work owned by another agent, consider an explicit agent handoff. Multiple agents are useful when independent work can be verified and combined; adding them to a simple sequence introduces coordination overhead.
Questions
Questions about Managed Agent Runtimes.
Is the Agents API generally available?
The release changelog describes it as public beta as of 19 September 2026. Confirm current access, limits and contractual requirements before a production commitment.
Can the agent write into our business systems?
Only through the tools and permissions your integration provides. Our proposed pilot starts with read-only investigation and introduces writes after explicit acceptance criteria and approval rules are tested.
Does a managed runtime replace evaluation?
No. Evaluate the full workflow against representative cases, including missing data, tool failures and a user changing the task midway through.
