Quietly, one September release changed how AI agents get built: OpenAI’s Agents API entered public beta — a managed agent harness reportedly priced around $0.03 per 20 minutes of agent runtime. Instead of writing the loop that plans, calls tools, reads results and retries, you describe the agent and rent the loop. Every serious platform now has a flavour of this. Here’s the concept, the trade-off, and why students should build the loop themselves once before renting it forever.
The three layers of agent-building in 2026
- DIY loop: you write the while-loop yourself — call model, parse tool request, execute, feed back, repeat (our from-scratch tutorial). Maximum understanding, maximum maintenance.
- Framework harness: libraries like LangGraph structure the loop for you — you still host and run everything.
- Managed harness (the new layer): the provider hosts the loop, the state, often a sandboxed workspace for the agent’s file and code operations; you send a task and tools, you get events and results. That’s what the Agents API sells — and it mirrors the sandboxed-VM thinking behind consumer agents like Meta’s Muse.
What “managed” buys you — and costs you
Buys: no loop code to debug, provider-maintained sandboxing (the hard security part), persistence across long tasks, and metered pricing that turns agent time into a utility bill. Costs: vendor lock-in (your agent’s shape follows their harness), less visibility into failures, per-minute costs that beat DIY only until your volume grows, and the classic platform risk — betas change. The professional pattern emerging everywhere: prototype on managed, understand via DIY, deploy on whichever the economics favour — the same build-vs-buy judgment every multi-agent design eventually faces.
Why students should still write the loop once
Interviews in 2026 have caught up with agents, and the differentiating question is no longer “have you used an agent?” — it’s “what does the loop do when a tool call fails?” Students who have written retries, timeouts, context-window trimming and tool-permission checks by hand answer in concrete detail; students who only called a managed API answer in brochure language. One weekend with the DIY tutorial earns that depth permanently. Then, having earned it, use managed harnesses freely — that’s not cheating, that’s engineering: you understand what you’re renting.
A sensible first managed-agent project
Pick a task with clear tools and a verifiable output — a placement-drive tracker (reads college notice text you paste, extracts companies/dates/eligibility into a sheet, flags conflicts with your timeline) is ideal: multi-step, tool-using, low-stakes when it errs, and demo-friendly. Build it DIY first, port it to a managed harness second, and write down the differences you felt — that half-page of notes is a better interview artifact than either version of the code. Agent infrastructure is becoming plumbing; the students who know where the pipes run will be the ones trusted to build with it.