Guides

AI teammates: how to brief an agent that runs all week

Perplexity and AWS just launched AI teammates that work for days at a time. This four-part brief decides what they do, when they ask and what they can touch.

S

SelfAgentic Team

7 min read
A hand writes a short to-do list in a notebook next to a laptop and a pair of glasses.

Perplexity launched Teammates on November 6: AI agents that each get a role, a standing brief and rules for when to involve you, then work in the background until something needs your decision. A day earlier, AWS shipped Kiro bots, persistent agents with their own name, notes, workspace and permissions. Both launches treat AI teammates as something you brief once and supervise, where a chatbot is something you prompt every time.

That raises the stakes on the brief. A weak prompt costs you one bad answer. A weak brief for an agent that runs all week costs you a week of slightly wrong work, or a decision made that should have been yours. A good brief has four parts: a role, a standing job with a clear definition of done, rules for when to stop and ask, and the list of tools the agent may touch. Three of the four match what Perplexity asks for when you set up a teammate, and the fourth is how Kiro scopes each bot.

What are AI teammates, and what did Perplexity and AWS ship?

An AI teammate is an agent that owns an ongoing job instead of answering one request. It keeps a persistent thread or memory, runs on a schedule or in the background, and comes back to you when it reaches a decision it isn't allowed to make.

Perplexity's Teammates announcement describes agents that plan their own tasks, bring in more agents when the workload grows, and keep going until the result is ready for your review. The examples on that page are useful because they spell out escalation rules. An Ads Manager teammate asks before increasing campaign spend. An On-call Engineer teammate pages you only when a fix needs approval or when it can't isolate the cause within 15 minutes. Each teammate keeps "a single thread that never closes." Engadget's launch coverage notes it's limited to Max subscribers, at $200 a month.

AWS's version is aimed at developers. VentureBeat's report on Kiro bots says each bot has a name, a continuing conversation, its own notes and a dedicated workspace. Permissions differ by bot: a research bot can inspect files without editing them, a deployment bot asks for approval before making changes, and a review bot can comment but can't approve or merge code. Kiro's bot documentation adds that each bot's memory and workspace are isolated from other bots.

VentureBeat also points out that AWS published no benchmarks or error rates with the launch. For now these are sensible design choices from two large vendors that still have to prove themselves in daily use.

How to write a brief for AI teammates

Think of the brief as an AI agent job description that fits on one screen. It has four parts.

  1. Role. One sentence naming the job and who it serves. "Researches new inbound leads for the founder" is a role. "Helps with sales" isn't, because the agent can't tell from it what to skip.
  2. Standing job and definition of done. What it does each time it runs, how often, and what a finished output looks like: a filled row, a draft in a folder, a three-line summary. If you can't describe done, the agent can't either, and it will keep working (and spending) past the point where you'd have stopped.
  3. Escalation rules. The specific moments it must stop and ask. Write them as triggers you could check: "before anything is sent to a customer", "if spend would go up", "if you can't find an answer after 15 minutes". Perplexity's on-call example is a good template because it has a type trigger (needs approval) and a time trigger (15 minutes).
  4. Tool access. The exact tools the role needs, and whether it can read, write or publish in each. A researcher needs to read your lead list. It has no reason to post on LinkedIn.

The fourth part is where products differ most. Perplexity's page says teammates use the skills and plugins you've installed, under the same permissions you have, and IT Brief's write-up adds that project permissions still apply in shared projects. Kiro sets permissions per bot. Inherited access is quicker to set up. Per-role access means a confused agent can do less damage. If your platform only offers inherited access, your escalation rules have to carry more of the load.

SelfAgentic took the per-role route on purpose. Connecting a tool and authorising a particular agent to use it are separate decisions, and an agent without a grant for a tool can't call it. The security page explains how those grants work.

A worked example: a lead research teammate for a three-person team

Say a three-person B2B startup gets 30 to 40 inbound demo requests a week, and the founder spends Monday mornings looking up each company. They brief one teammate like this:

  • Role: researches every new inbound demo request for the founder.
  • Standing job: each weekday at 7am, for every new request, find company size, what they sell and one recent public update, then write those into the lead sheet with a source link for each. Done means every new row has all three fields and a link.
  • Escalation: if the company looks like a competitor or an existing customer, flag the row and do nothing else. If no reliable source turns up after ten minutes of searching, mark the row "needs a human" and move on.
  • Tools: read the sheet where demo requests land, read and write the lead sheet, search the web. Nothing that sends email or posts anywhere.

On Monday the founder opens the sheet and finds, say, 34 rows filled and 3 marked "needs a human". Those three rows deserve the review time, because they show where the brief met something it didn't cover. If two of them turn out to be agencies booking on behalf of clients, that becomes a new line in the brief, and next week the agent handles it.

Nothing in that setup lets the agent contact anyone. A second teammate could draft outreach for the founder to approve, with its own tools. That split between a researcher and a drafter is manager and specialist delegation in practice. On SelfAgentic, a manager agent hands pieces of work to specialist agents this way, and outward-facing actions such as LinkedIn posts or Meta Ads campaign changes wait in an Approvals inbox by default on scheduled runs.

Where AI teammates still fall short

Briefs drift. The job changes faster than anyone updates the brief. VentureBeat reports that Kiro lets you inspect, revise or remove what a bot remembers. Whatever tool you use, put a monthly reminder on your calendar to reread each brief and what the agent has stored.

Approval requests pile up. An agent that asks too often trains you to click approve without reading. If you approve the same request ten times in a row unchanged, turn it into a written rule.

Costs are harder to predict. Perplexity says teammate usage draws on the credits already in the Max plan once a promotional period ends, and VentureBeat's Kiro piece warns that "a credit is not a completed task." An agent that runs all week can spend all week, so set a budget per teammate or at least know where to check spend.

Parallel agents multiply mistakes along with output. When one teammate brings in more agents, a vague brief gets copied to all of them.

FAQ about AI teammates

What's the difference between an AI teammate and an AI assistant?

An assistant answers when you ask. A teammate owns a recurring job, keeps its context between runs and contacts you when it needs a decision. In practice you review its output on its schedule instead of watching each step.

Do AI teammates need their own permissions?

Ideally, yes. Some products let teammates inherit your access, which is faster to set up. Giving each role only the tools it needs limits what a mistaken or manipulated agent can reach, and that matters more the longer it runs unattended.

How many AI teammates should a small team start with?

Start with one, on a job you already do every week and can check quickly. Add a second after you've rewritten the first one's brief at least once based on what it got wrong.

What should an AI teammate never do without approval?

Anything that reaches people outside your team or spends money. That covers sending email, publishing posts and changing ad budgets.

Where to start this week

Pick one job you do every week. Write its four-part brief in a doc before you open any tool, then run it for two weeks with every outward action set to need approval. The rows marked "needs a human" will tell you what the brief is missing. If you want to try it with agents that each get their own tool grants and remember context between runs, you can start with SelfAgentic.

Keep reading

Ready to build your AI team?

See how a team of agents handles real work, and checks with you before anything goes out.

Free to start. No credit card required.