How should agent teams work? Modeling human organization patterns 🧠

[EngineerProjects]

The question at the heart of Seshat

Individual agents are increasingly capable. They can reason, use tools, write code, search the web, and produce real deliverables. That problem is largely solved.

The frontier has moved. The hard question now is:

How do multiple autonomous agents work together in a coherent, observable, and useful way?

---

Our starting hypothesis

Most multi-agent systems today rely on pipelines, handoffs, or orchestration graphs. These work for structured workflows, but they feel brittle and opaque for anything complex and open-ended.

We think the better model is the human organization not as a metaphor, but as an actual design target.

Real human teams do not communicate through function calls. They:

• 📬 Send messages and reports through explicit channels (email, Slack, meetings)
• 🎯 Have roles and responsibilities, not just capabilities
• 📋 Assign, claim, and escalate tasks
• 🗂️ Maintain a shared mission memory: objectives, decisions, constraints
• 📅 Organize reviews, syncs, and handoffs deliberately
• 🔍 Produce observable outputs — reports, summaries, changelogs not just internal state

The idea is to study how humans actually coordinate in teams in companies, research groups, open-source projects, or any collaborative setting and reproduce those patterns with autonomous agents.

---

What we want to know from you

This is an open design question and we want real input before we build.

Some specific things we are thinking about:

1. Communication primitives : Should agents communicate via message queues? Shared whiteboards? Something else? What is the right abstraction?

2. Role vs. capability : Is it enough to give an agent a system prompt with a role, or do we need first-class role objects with permissions, responsibilities, and reporting lines?

3. Mission memory : How do you keep a team of agents aligned over time without sharing everything? What belongs in shared memory vs. individual context?

4. Meetings and synchronization : Does it make sense to model agent "meetings" structured moments where agents exchange state and make joint decisions? Or is async-first always better?

5. Observable work : How do you make agent team work legible to the humans supervising it? Dashboards? Structured reports? Activity feeds?

6. Human-in-the-loop : At what granularity should a human be able to intervene, redirect, or override an agent team mid-mission?

---

What organizational patterns are worth modeling?

We are also curious about which real-world organizational structures are the most interesting to reproduce:

• Classic hierarchies (manager, specialists, reviewer)?
• Flat teams with explicit task markets (anyone can claim a task)?
• Committee-based decision making (multiple agents vote or review before acting)?
• Newsroom-style (editor, reporters, fact-checker, publisher)?
• Something else entirely?

---

Share your thoughts

Drop a comment below. There are no wrong answers, we are at the design stage and every perspective helps shape what gets built.

If you have seen a multi-agent coordination pattern that worked well (or badly), we want to hear about it. If you have a concrete use case in mind, describe it. If you think the whole human-organization analogy is wrong, make the case.

This is a genuinely open question. 👇