Disclosure: I work at Conduyt, a CRM company, and I use our product as the worked example near the end. The rules apply to any system that puts people and agents in the same channel.
Most teams adopted AI agents one laptop at a time. One person runs Claude Code in a terminal. Someone else prefers Codex. Each agent is useful to its owner and invisible to everyone else, so when a colleague needs the same answer, they start from scratch or ask for a paste.
The obvious fix is to put the agents where the team already talks: a shared channel where anyone can ask an agent a question and everyone can see the answer. That part is easy to build. What's harder is that a shared channel changes who an agent is acting for, what it can reach, and who is watching it.
Here are the four rules I would want settled before any team does this, whichever tool they use.
Rule 1: An agent never sees more than its owner
In a private terminal, this barely comes up. In a shared channel, a question can come from anyone, while the answer may come from an agent running with someone else's access.
The OWASP Top 10 for LLM applications names this risk Excessive Agency and traces it to three root causes: too much functionality, too many permissions, and too much autonomy. A shared agent with a broad service account has all three.
The simple rule is that each agent acts with one named person's permissions, and never more. If that person can't see a record, neither can their agent. This also gives you a clear answer to the first question anyone asks after a mistake: whose agent was it, and what was it allowed to do?
It doesn't solve everything. As Sebastian Martinez Torregrosa argues in Two Mandates, One Walk, two agents that are each properly bounded can still reach something together that neither could reach alone. So if agents can hand work to each other, review the handoff as well as each agent.
Rule 2: Nothing leaves in a person's name without that person
Reading data is low risk. Sending an email to a customer under someone's name is not, and an agent drafting one in a busy channel makes the mistake easier to miss.
OWASP's mitigation is human-in-the-loop approval for high-impact actions. But the design matters. In Human Approval Is a Protocol, Not a Button, Raju Dandigam walks through a good failure case: a manager approves a $250 payment, the agent replans, and $2,500 goes out because the system only checked a yes/no flag.
So when you test any system like this, try to change the content after approval. Does the approval still hold? It shouldn't.
Rule 3: Agent conversations need a stopping condition
Once agents can @mention each other, they can also keep replying to each other. Nobody has to intend that. Two helpful agents and an ambiguous request will do it.
Anthropic's write-up on its multi-agent research system is a useful reality check. It describes early agents spawning 50 subagents for simple queries and distracting each other with excessive updates. Its guide to building effective agents recommends stopping conditions, such as a maximum number of iterations, to maintain control.
In a shared channel, I'd want a hard limit on consecutive agent turns, after which the thread waits for a person. The exact number matters less than the fact that it's enforced outside the model. A prompt asking agents to please stop isn't a control.
Rule 4: The work stays where the team can see it
If an agent's output lives only in one person's session, you've saved that person time and left the handoff as broken as it was. The point of sharing is that the question, the answer and the record it refers to all sit in the same thread, so the next person can pick it up.
This also makes supervision cheap. People catch bad answers when they can see them, and in a shared channel they can, without anyone building a review queue.
A worked example
Here is how we applied this at Conduyt. Team Chat sits inside the CRM, and AI teammates join the same channels as people. Today that means Claude Code, Codex or Grok Bot.
- Permissions: each agent acts with its owner's permissions and never more. Claude Code and Codex stay on their owner's computer and pair through a small listener with a single-use code that expires after 15 minutes. Grok Bot connects through its own connector.
- Sending: anything an agent wants to send under a person's name waits for that person to approve it, edit it first or decline it.
- Stopping: after six AI turns in a row, the thread pauses until a person weighs in.
- Visibility: the conversation sits beside the deals and contacts it refers to.
There are real limits. The list of supported agents is short. AI teammates are on our Professional plan, not the entry plan. And six turns is a judgment call, not a proven number. I posted a full transcript of three agents working one thread, with the engineering decisions behind it, in an earlier write-up. If you're comparing options, the wider field of AI CRMs takes very different approaches, and some teams will be better served by something simpler.
How to test any of this
Skip the scripted demo. Run one real task with two agents belonging to two different people:
- Log in as someone with narrow permissions and check what their agent can see.
- Have an agent draft a customer message, approve it, then try to change it.
- Give two agents an ambiguous request and see what stops them.
- Ask a third person to pick up the thread and finish the job.
Each step tests one of the rules, and you'll learn more from them than from any feature list. If a product can't pass the first two, I wouldn't give its agents access to customer data, however good the model is.
Jordan Tate is VP of Customer Relations at Conduyt, a flat-price CRM for small sales teams.
Top comments (0)