OpenAI
Managed execution, browser actions, and finished artifacts
The relevant change is the unit of delegation: an agent can work across tools toward a reviewable artifact. The teaching question is which intellectual decisions remain visible.
People and decisions
Questions to examine before considering a course use:
- Which choices require disciplinary judgment, even when an agent can produce a finished lesson?
- Does a proposed saving in preparation time survive review, correction and the needs of students with different access?
What the company has described
Agents API enters public beta
Managed sessions support continuing work, tools, and execution environments.
Public beta is a product status, not an instructional validation.
Source and inspection scopeHosted browser computer use
Agents API adds a browser operated through a managed session.
Website access approval does not guarantee approval before each consequential action; see computer-use documentation.
Source and inspection scopeDocument skills in hosted sandboxes
Hosted sandboxes add skills for spreadsheets, presentations, and Word documents.
A finished file still needs content and accessibility review.
Source and inspection scope
What this means for a psychology instructor
The following is instructional analysis. Classroom examples are original, unevaluated proposals; they do not describe tested benefits of this company’s products.
From an answer to a course artifact
Instructional interpretation: a psychology instructor should evaluate a delegated lesson as a sequence of decisions. Start with a narrow objective, such as helping students distinguish random assignment from random sampling. Specify the intended misconception, an example population, and the evidence students must produce. A proposed agent task could assemble a short reading, construct two contrasting study descriptions, and return a draft slide sequence with a source table. The worthwhile output is the connection between these parts. An attractive presentation that quietly swaps the two sampling concepts would fail the task even if every file opens correctly.
Where agency matters
In this original example, a fixed workflow would always retrieve three documents, generate six slides, and export them. An agentic design would be allowed to notice a missing definition, inspect another permitted source, revise the example, and explain why its plan changed. Those additional decisions create something worth inspecting. The instructor can ask whether a change corrected the conceptual problem or merely added plausible prose. The number of tools used is a poor success criterion. A useful acceptance record would instead show which claims were checked, which examples were rejected, and which course decisions were left unresolved.
What to measure before expanding
A small local pilot could compare preparation time and revision burden for the same lesson prepared with and without the proposed agent task. Keep the materials and acceptance criteria fixed. Record conceptual errors, source mismatches, inaccessible diagrams, and the instructor time needed to find them. Test students separately with a new study description that does not repeat the prepared wording. That is an unevaluated proposal for assessing a workflow, not a claim about OpenAI improving learning. When browser actions or course systems enter the task, separate drafting from release. A reviewable lesson file and a published course announcement have different consequences. Permission to consult a course website should not silently become permission to alter it.
Decisions to examine
- Proposed pilot: delegate a bounded lesson artifact, then inspect conceptual alignment and revision time.
- Use an unassisted transfer question to evaluate student understanding separately from material quality.
- Decide which actions may create drafts and which require a separate release decision.
Open questions
- Which errors remain invisible to an instructor who is not already expert in the topic?
- Does saved preparation time survive source checking and accessibility review?
- What workflow state and student material would be retained under the institution’s actual configuration?
Original sources and limits
Compare other companiesHow agentic systems workTeaching implications