Why do multi-agent AI systems fail?

Multi-agent systems fail at coordination, not intelligence. The models are usually good enough; what breaks is the hand-off. Context dies between agents, nobody owns the next step, and work falls between roles.

What is a coordination layer for AI agents?

A coordination layer is the shared infrastructure that lets one employee's AI agent work directly with another's, carrying identity, scoped context, commitments and evidence between them, underneath the tools the company already uses.

Should every employee have their own AI agent?

Most already do. The question that matters is whether those agents can reach each other. One agent per employee is the default state of a 2026 company; the missing piece is the layer between them.

How do you stop an AI agent acting beyond its authority?

Separate what an agent may see from what it may commit to, and enforce the split outside the model. Read scope, write scope and commit scope are different planes, with human gates on anything irreversible: a merge, a customer message, a payment, a production change.

What is AI context lock-in?

Context lock-in is what happens when a model vendor becomes the coworker that remembers, routes and executes. The company ends up renting its own operating memory, and switching models means losing everything the organization has learned.

How many AI agents does a company need?

Agent count is the wrong metric. Fifty agents is not one company. What matters is how many hand-offs happen without a human ferrying context, and that number should be rising while headcount stays flat.

Why sell an agent product to the manager, not the developer?

Because the manager is the integration. Developers already have an agent they like and resist anything new in the editor; the engineering manager or PM is the person pasting between those agents, chasing status and being surprised on Friday. They are the one who is actually paying for the gap.

What does a coordination layer have to beat?

Not Slack, not Linear, not a standup bot. The real competitor in every scenario is 'I ping the person in Slack and their Claude answers in ten seconds.' A coordination layer earns its place on what is left after that: the question cannot be silently dropped, the other person's agent answers while they sleep, and there is a record instead of a vibe.

How many companies actually use AI agents?

About half of large enterprises are deploying agents, and roughly one in five has them working across a workflow. KPMG's Q2 2026 survey of 204 US leaders at companies above $1B found 53% deploying agents and 18% orchestrating multiple agents across workflows. Deloitte puts scaled cross-functional multi-agent adoption at 15%. Adoption is real. Coordination is not.

Field notes · Pantheon

Field notes

The questions we had to settle first.

One question per note. The answers became Pantheon.

Note 09
About half of large enterprises are deploying agents, and roughly one in five has them working across a workflow. KPMG's Q2 2026 survey of 204 US leaders at companies above $1B found 53% deploying agents and 18% orchestrating multiple agents across workflows. Deloitte puts scaled cross-functional multi-agent adoption at 15%. Adoption is real. Coordination is not.
Subject: Adoption
3 Sep 20264 min read
Note 08
Not Slack, not Linear, not a standup bot. The real competitor in every scenario is 'I ping the person in Slack and their Claude answers in ten seconds.' A coordination layer earns its place on what is left after that: the question cannot be silently dropped, the other person's agent answers while they sleep, and there is a record instead of a vibe.
Subject: Positioning
2 Sep 20263 min read
Note 07
Because the manager is the integration. Developers already have an agent they like and resist anything new in the editor; the engineering manager or PM is the person pasting between those agents, chasing status and being surprised on Friday. They are the one who is actually paying for the gap.
Subject: Strategy
2 Sep 20263 min read
Note 06
Agent count is the wrong metric. Fifty agents is not one company. What matters is how many hand-offs happen without a human ferrying context, and that number should be rising while headcount stays flat.
Subject: Metrics
1 Sep 20262 min read
Note 05
Context lock-in is what happens when a model vendor becomes the coworker that remembers, routes and executes. The company ends up renting its own operating memory, and switching models means losing everything the organization has learned.
Subject: Strategy
26 Aug 20263 min read
Note 04
Separate what an agent may see from what it may commit to, and enforce the split outside the model. Read scope, write scope and commit scope are different planes, with human gates on anything irreversible: a merge, a customer message, a payment, a production change.
Subject: Security
24 Aug 20263 min read
Note 03
Most already do. The question that matters is whether those agents can reach each other. One agent per employee is the default state of a 2026 company; the missing piece is the layer between them.
Subject: Adoption
21 Aug 20263 min read
Note 02
A coordination layer is the shared infrastructure that lets one employee's AI agent work directly with another's, carrying identity, scoped context, commitments and evidence between them, underneath the tools the company already uses.
Subject: Definition
18 Aug 20263 min read
Note 01
Multi-agent systems fail at coordination, not intelligence. The models are usually good enough; what breaks is the hand-off. Context dies between agents, nobody owns the next step, and work falls between roles.
Subject: Coordination
12 Aug 20263 min read

9 notes · newest first · field notes from building the coordination layerThe companies described are composites, not customers.