Anthropic
Persistent agent sessions, dynamic workflows, and explicit execution controls
Longer-running delegation makes intermediate decisions and retained context part of the teaching design. More agents do not create independent evidence.
People and decisions
Questions to examine before considering a course use:
- Which teaching decisions should remain with an instructor even when a session can continue working?
- Who can inspect, correct or stop an ongoing task, and who has time to do so?
What the company has described
Managed-agent architecture described
Anthropic describes separating agent control, persistent history, and execution.
Engineering account, not an independent classroom evaluation.
Source and inspection scopeDynamic workflows in Managed Agents
In beta, an agent can write a phased program coordinating many agents and combining results.
A generated plan may reproduce the same mistaken premise across tasks.
Source and inspection scopeOrganization web restrictions apply to agent tools
Managed Agents web fetch follows organizational domain restrictions as well as agent-level settings.
Domain restrictions constrain access, not the truth of an allowed page.
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.
What could persist between preparation sessions
Instructional interpretation: a persistent course-preparation task needs a versioned course brief. Imagine revising a research-methods unit over several days. The brief records what students already know, the distinction between measurement reliability and validity, and the assignment that will provide evidence of understanding. A proposed agent might identify inconsistent examples across a handout and a rubric, then return a discrepancy list. The durable part should be the approved brief and the record of corrections. A previous conversational suggestion should not become course policy merely because it remains available in a session. The instructor needs to distinguish a retained instruction from a retained hypothesis.
Multiple agents, one acceptance decision
An original multi-agent exercise could divide a synthetic assessment into three reviews: conceptual alignment, clarity for novice readers, and the plausibility of distractors. Give each review the same explicit learning objective but a different question to answer. Then inspect disagreements before changing the task. Agreement among several agents would not show that a question measures the intended construct; all may inherit the same faulty definition. A human reviewer should ask for counterexamples, inspect the source passage, and test whether students could answer through superficial cues. The purpose of delegation is to expose more inspectable work, not to replace an assessment argument with a vote.
Control should follow the consequence
For a first pilot, use invented student responses and a copy of the course materials. Ask the proposed workflow to suggest feedback categories, then compare its suggestions with the instructor’s own coding. Examine false equivalences: a student who confuses correlation with causation needs different feedback from a student who recognizes the limitation but expresses it poorly. Avoid inferring dishonesty, ability, or diagnosis from prose. Persistent sessions also require an explicit institutional decision about what material is appropriate to retain. A pause or an access restriction can stop an operation, but it does not establish the educational quality of the feedback that was produced. Expand the task only after its errors and review burden are understood.
Decisions to examine
- Maintain an approved course brief separately from provisional suggestions.
- Use independent review questions and counterexamples instead of treating agent agreement as verification.
- Pilot feedback classification with synthetic responses before considering any student work.
Open questions
- Can reviewers reconstruct why a persistent task changed a course example?
- Do parallel reviews find different errors or repeat the same assumptions?
- Which retained history is needed, who can access it, and when should it be deleted?
Original sources and limits
Compare other companiesHow agentic systems workTeaching implications