Persistent AI Agents Shift the Job to Oversight

From prompt to standing assignment

Perplexity’s Teammates announcement points to a different unit of work for AI: not one answer, but an ongoing assignment. The company describes agents that remember project goals, task history and shared files across runs; monitor work as conditions change; and coordinate sessions or hand tasks to other teammates. It says people can inspect progress, redirect an agent or take over its desktop. The initial release is web-based for Perplexity Max subscribers, with iOS support described as coming soon. These are vendor-described capabilities, not an independent reliability test. S1

That changes the product question. A chat assistant is judged on its answer; a persistent agent also has to notice that something changed, retain the right context and know when not to act. Memory can prevent repetitive setup, but it can also preserve a stale assumption. Monitoring increases the opportunities to help—and the opportunities to make an unwanted change.

Set the boundary before adding memory

A Y Combinator Startup Library conversation with Gumloop’s founders offers a useful counterpoint to the hands-off pitch: people who understand a task should shape its automation, while IT controls data access, security and cost. S2 For developers, that means starting with a bounded contract: one accountable task owner, narrowly scoped connectors, explicit triggers, and a clear stop condition. Separate preparing a draft from sending it; require approval before publishing, deleting or changing production. Keep a record of what the agent observed, what it changed and what it handed back.

This is not an argument against autonomy. It is a way to make autonomy legible. A recurring low-impact task can be a good first assignment; a broad instruction like “keep the team on track” is hard to test and easy to over-authorize. Let the owner define what counts as done and which actions need a checkpoint before enabling unattended runs.

Make persistence earn its keep

The implementation details need a date label: Anthropic’s agent-building guide warns that much of its tooling landscape has changed since December 2024. Its durable caution is that agentic systems trade latency and cost for potential performance gains, so complexity should earn its place. S3

For each candidate workflow, compare a persistent agent with a scheduled script or fixed sequence. Measure completed tasks, safe refusals, human corrections, recovery time and operating cost—not just successful demos. Replay cases with stale context, changed priorities, missing permissions and duplicate triggers. Check that the agent can pause and explain what it needs when the task crosses a boundary.

The practical takeaway is simple: persistent memory is a capability, not a safety property. Start small, keep permissions narrow and expand only when repeated runs show that the agent can detect uncertainty, surface its actions and hand control back. The most useful teammate may be the one that knows when to stop and ask.