Systems of several agents fail in ways that repeat. This chapter goes through the most common ones and the piece of structure the model places in front of each. No piece is a guarantee (all of them lower the probability of a failure, none eliminates it), but together they change the risk profile considerably.
The handoff that loses context
The failure. One agent passes work to another, and along the way what mattered gets lost: a warning, a decision already made, the exact command that had to be run. The receiver starts from scratch or, worse, starts out convinced of something wrong.
What stops it. The work lives in a shared record with history (chapter 6), not in the head of whoever passes it on. The receiver reads the real state, not a summary. And whoever receives a request states what they understood before starting (chapter 11), so a bad handoff shows up in the first line.
Noisy context
The failure. A conversation accumulates so much material (failed attempts, long documents, back and forth) that the agent starts picking irrelevant details or forgetting what matters.
What stops it. Several pieces at once. Skills are read only when they are needed, instead of all being loaded all the time. Tasks and subagents run separately, and hand the main conversation a result, not their whole path. And since the work is saved outside, a conversation can be restarted without losing anything: each day starts from a note (chapter 6), and long work is given to a new conversation with a written request, or to one that starts from a targeted summary of what came before, and not to a copy that inherits the entire context of whoever opens it. A small, relevant context performs better than a large, noisy one.
Agents that agree with each other
The failure. An agent reviews the work of another that shares its same assumptions, and approves it. The review exists, but it reviews nothing.
What stops it. Subagents start without the context of whoever creates them (chapter 4). A reviewer that did not see the conversation does not inherit its assumptions, and that is why it is useful for looking with fresh eyes. One caveat applies: a reviewer without context can also be confidently wrong, and that is why its findings are verified before they are applied, especially when they are serious.
The endless chain
The failure. Two or more agents pass work back and forth, each waiting for the other to finish, and the chain never breaks (with the cost that implies).
What stops it. The traffic rule (each conversation talks only to its role's main conversation, and the main conversations cross paths through whoever coordinates, except for direct questions between leaders) reduces the possible paths. And a maximum number of consecutive turns without a person in between cuts any chain that gets loose. The brake does not depend on the agents noticing: it is in the structure.
The order disguised as data
The failure. A text arriving from outside (a customer's message, a web page, a file) contains instructions, and the agent follows them as if they came from someone authorized.
What stops it. The model treats all of that as data and never as an order: events arrive quoted with their origin, customers' messages are read as text from someone outside, memories are presented as notes. And even if an agent were convinced, the access ceiling does not move with what the agent reads, and the credentials are never within its reach.
The signal nobody hears
The failure. One conversation waits for something from another (an answer, a finished task) through a shared board, and the signal never arrives: the filter it waits with does not match what gets posted, the question reaches a conversation that has already ended, or the task completes without notice. The wait lasts until it times out.
What stops it. A shared board is useful for seeing who owes what, but it is not enough on its own: it needs a mechanism that wakes whoever is waiting. Each board event that matters to a conversation becomes a delivery for it, just like a message, and finishing a task also sends notice. Without that bridge, the board is looked at; with it, the board moves the work.
Invisible work
The failure. The agents do things that nobody in the area knows are being done, and when something goes wrong, nobody can reconstruct what happened.
What stops it. The role's people see what its routines and tasks are doing, and those who direct the role can pause them. Every change stays in a record with who and when. And every request comes in and goes out through known places.
What structure does not stop
To be fair to the reader, it is worth naming what none of these pieces solves. A role with bad instructions will do its work badly, with all the structure in the world. Each role's judgment depends on how it was written, and that work still belongs to people (chapter 17). Structure reduces coordination failures; judgment failures are reduced by directing well.
In short, each typical failure of agent systems has, in this model, a piece of structure that stops it, and no piece replaces careful direction. The next chapter looks at how a business factory improves over time.