Public transcript

How to Build Team Agents

By: The AI Daily Brief: Artificial Intelligence News and Analysis
Length: 42 min
Published: 29 September 2026
Language: English
Text from: the audio
Transcribed: 30 September 2026

Read the full transcript Open on Spotify

Notes in English.

Summary

This Operator's Cut episode of the AI Daily Brief features Nufar Gaspar explaining how to build team agents - agents shared across a team with shared knowledge, memory, and configuration. She argues that companies naturally evolve from individual agents to team-level agents, identifies four archetypes of team agents, and walks through the five key design decisions: what it does, where it lives, what it knows, what it can touch, and how to run it. She closes with the advice that the heavy lifting is always configuration and knowledge curation, not tool selection.

Key points

Companies progress through three steps: everyone builds their own agents, then gets stuck in an agent sprawl, then merges agents into team-level agents that are maintained by people and refined over time.
Not every agent should be shared - there is a spectrum from private agents, to shared knowledge with private agents, to full team agents.
The four types of team agents are the expert agent, the common work agent, the bridge agent, and the chief of staff agent.
Three signs you should not build a team agent yet: taste beats standards, nobody can own the knowledge, and it complicates more than it simplifies.
The five design decisions are what it does, where it lives, what it knows, what it can touch, and how you run it.
For access, the safest option is for the agent to act as whoever is asking, rather than using its own account or one person's login.
The expert agent is the most common and lowest-hanging fruit starting point, while the bridge agent is the most attractive conceptually but hardest to execute.
The heavy lifting is always configuration and knowledge curation - select the tool closest to your existing ecosystem and invest in agreeing on ground truth, dos and don'ts, and use cases.

Questions it answers

00:00Why team agents matter

Why should you care about team agents right now?

Most work happens in teams between people, not just individually, but most agents have been solo affairs. AI-forward companies are moving from individual agent sprawl to merged team-level agents, and this pattern is happening publicly at companies like Avery, Sierra, and Shopify.

Companies progress from individual agents to agent sprawl to merged team agents.
Work happens at the intersection between people, which is where team agents live.

08:06The sharing spectrum

Should every agent be shared with your team?

No, there is a spectrum with three settings: private agents with your own taste and access, shared knowledge where each person maintains their own agent but points to a shared knowledge base, and full team agents where one agent serves many people. The right choice depends on whether taste or standards matter more for the use case.

If colleagues keep asking to borrow your private agent, consider sharing it.
Shared knowledge is often the easiest and right place to start.

15:39Four archetypes of team agents

What types of team agents can you build?

There are four kinds: the expert agent (captures one person's know-how), the common work agent (unifies similar recurring work), the bridge agent (handles work flowing between roles where nobody can do it alone), and the chief of staff agent (owns day-to-day team operations and onboarding).

The expert agent is the lowest-hanging fruit and most common starting point.
The bridge agent is conceptually most attractive but hardest to execute.
Knowing your archetype helps you identify use cases and know what to pay attention to.

18:47When not to build

What are the signs that a team agent is the wrong move?

Three signs: when taste beats standards (people need different answers and private judgment), when nobody can own the knowledge (the agent will drift within weeks), and when it complicates more than it simplifies due to conflicting needs or tangled permissions. Sensitive data and high stakes are not reasons against building - they are design questions.

Sort out ownership before building, otherwise the agent drifts quickly.
Sensitive data shapes how you build, not whether you build.

19:59What it does and where it lives

How do you scope a team agent and choose where to host it?

Define who it serves by role, its broad area of work, and a don'ts list including permissions and data handling. For hosting, options range from a simple shared folder with existing tools, to vendor-hosted agents like Claude Tag or ChatGPT workspace agents, to self-hosted open-source agents. Pick the simplest option two people will actually use this week.

The don'ts list is the part people often skip.
Decide who can see conversations and where learning is stored - trust is gained or lost here.
Start with narrower scope (reading and drafting) and expand only after trust is earned.

26:57What it knows and what it can touch

How do you manage knowledge and access permissions for a team agent?

Knowledge goes through four stages: collect from all sources, refine by surfacing contradictions, approve with each piece signed off by its owner, and maintain with a schedule for agent-proposed updates reviewed by a person. For access, the safest option is acting as whoever is asking; sensitive answers must go privately to the asker, not shared channels.

The knowledge conversation is worth having even if you never ship the agent.
Claude in Slack currently doesn't consider who else is in the channel when surfacing answers.
Keep records of who asked for what when the agent works under its own account.

33:07How to run it

What makes a team agent live beyond its first week?

Four things: one owner who maintains priorities and resolves conflicts, clear rules of engagement telling people how to work with it, putting decisions in places the agent can see, and continuous monitoring with a small pilot group and rerunnable test questions.

People trust agents more when they know what it's learning and the learning process.
If decisions happen in private messages and hallway conversations, the agent never hears about them.

37:57Tool choice and predictions

How much should you invest now versus waiting for better tools?

The heavy lifting is always configuration and knowledge curation, not tool selection. Pick the tool closest to your existing ecosystem, focus on agreeing on ground truth and dos and don'ts, and you will be ready when better tools arrive. Your moat is in what the tool companies cannot tap into.

More formalization of shared spaces, permissions, and ownership is coming.
Building your own version of Claude Tag now may not be worth the energy.
Your moat is in the knowledge and business decisions, not the tools.

Notes

Why Team Agents Matter

Work often happens between people, but most agents have been solo affairs covering only individual work.
Companies progress through three steps: (1) everyone builds their own agents, (2) agent sprawl emerges with overlapping work and inconsistent pictures of the company, (3) the most AI-forward companies merge agents into team-level agents maintained by people and refined over time.
Examples: Avery gave every employee an agent early in the year then moved to shared team agents by May; Sierra merged many specialist agents into one; Shopify has internal agents.

The Spectrum

Private agent: yours alone, with your taste and access (e.g., a personal social media agent).
Shared knowledge: team maintains one body of knowledge (e.g., ideal customer profile, messaging) while individuals keep their own agents pointed at it.
Team agent: one agent many people work with, with shared knowledge, memory, and configuration.
If colleagues keep asking to borrow your private agent, that is a sign to consider sharing it.

Four Archetypes

1. Expert agent - one person's or a small team's know-how made available to everyone; requires expert involvement from day one; creates vacation relief but also job insecurity concerns.

2. Common work agent - similar recurring work done one shared way; recognized when three people have each built their own version; requires agreeing on standards.

3. Bridge agent - work that flows between roles where nobody can do it alone; requires knowledge from each function and careful attention to permissions.

4. Chief of staff agent - owns team operationalization of day-to-day work, status, onboarding; requires clearly defining what it can learn and how.

When Not to Build

When taste beats standards - people need different answers and private judgment.
When nobody can own the knowledge - the agent will drift within weeks.
When it complicates more than it simplifies - conflicting needs, tangled permissions, endless coordination.
Sensitive data and high stakes are design questions, not reasons against building.

Five Design Decisions

1. What it does: who it serves by role, a broad area of work, a don'ts list (never makes commitments, never settles disagreements, never carries info between private spaces), start narrow with reading and drafting.

2. Where it lives: shared folder with existing tools, vendor-hosted ready-made agent (Claude Tag, ChatGPT workspace agent, Copilot, Notion), or self-hosted (OpenClaw, Hermes, custom harnesses). Decide who sees conversations and where learning is stored. Pick the simplest option two people will actually use this week.

3. What it knows: collect (aggregate from all sources), refine (merge, surface contradictions, date everything, keep out what should never be shared), approve (each piece signed off by its owner), maintain (agent proposes updates, person reviews and approves).

4. What it can touch: decide whose access it uses - act as the asker (safest), its own account, or one person's login (read-only only). Decide who can ask it. Decide where answers land - sensitive answers must go privately to the asker, not a shared channel. Keep records of who asked for what.

5. How to run it: one owner, clear rules of engagement, put decisions where the agent can see them, keep watching with a small pilot group and test questions.

Closing Insights

The expert agent is the lowest-hanging fruit and most common starting point; the bridge agent is conceptually most attractive but hardest to execute.
The heavy lifting is always configuration and knowledge curation. Pick the tool closest to your existing ecosystem and focus on agreeing on ground truth - your moat is in what the tool companies cannot tap into.

Transcribed automatically from the audio. Summary and notes written by AI from the transcript, so check anything important against the recording.