Someone on a customer's team asked one of our agents a question in a busy Slack channel. The agent had context from a DM a colleague had sent it the week before. So it did what a good coworker does. It said that colleague had been looking into it and was the right person to ask.

That's a small reply with a lot underneath it. An always-on teammate carries context from everyone it works with, including conversations the person asking was never part of. What was said in private stays private, and the agent never repeats it. The judgment is in everything around that line. Sometimes the helpful move is pointing someone to the colleague who knows. Sometimes even that says too much, like when the DM was about a surprise birthday plan for the person asking, and the right move is to say nothing.

No benchmark will ever score a moment like that, but it's exactly why that colleague keeps telling the agent things. The gap between doing the task and being trusted with the next one is most of what we think about now. It's also the biggest reason we build our agents on Claude Opus.

‍

Always-on is a different test

Workforce agents like Ace are digital teammates. They have their own accounts in Slack, email, and the rest of a company's tools, take requests from the people they work with, and stay available to the team all day. (I wrote about why that identity matters here.)

The model is a small part of that, and it's the part everything else stands on. Nobody tours a house to admire the foundation, but the foundation decides how tall you can build and whether what you put on top stays up. Our floor and our ceiling both start there.

A chat assistant gets graded on one answer. An always-on teammate is in dozens of conversations at once, with people who talk to each other about it, and most of what it does is visible to someone. A hundred good small calls earn a little trust. One bad visible one costs a lot.

Most of us have worked with someone brilliant who was still hard to work with. Great at the work, exhausting to be around. Being good at a job and being good to work with were always two different things, and agents are no exception. For an agent, that comes down to small calls.

‍

The small calls

Three come up constantly.

Is this mine? Being @mentioned in a group channel doesn't mean you're the one being asked. Neither does "can you guys..." Two of our agents on the same team learned this the funny way. They kept jumping on the same requests, five collisions in one afternoon, until they landed where every human team eventually does: say "I've got this," wait a beat, then start. Two agents racing to be helpful is the most wholesome bug we've ever shipped.

Is now the time? An agent that's always available can answer at 2am, post an update every twenty minutes, and nudge a stalled request every hour. That's exhausting to work next to. So our agents pace themselves the way a good colleague does: a quick 👀 when they can't get to something yet, an update when something actually changed, and a follow-up after a reasonable gap.

Should I speak up unprompted? Lucas wrote in Seeing the Forest about when an agent should stop and ask. This is the other direction. An agent in that many threads notices things: two people solving the same problem in different channels or a deadline quietly about to slip. Flagging those is some of the most valuable work it does. Flagging everything is noise, and a teammate who cries wolf gets muted.

‍

Manners aren't permission

Social rules live in a custom per-agent handbook. Hard boundaries live in the architecture: plain old software that sits underneath every small call.

The system labels every conversation as public, private, a DM, or a group DM, so the agent always knows which one it's in. It can't post in a channel it isn't a member of. It can't reply to a thread with new messages it hasn't read, which rules out a whole category of awkward replies to conversations that already moved on. It can't edit a task that changed since it last looked.

The software makes sure the agent works from current facts, and judgment covers everything above that.

‍

Benchmarks test strangers

Every model launch comes with a wall of benchmark charts, and we read them like everyone else. But a benchmark is a single episode with a stranger. The task starts, the task ends, and nobody remembers how it went.

Work doesn't reset. The agent sees the same people tomorrow and the day after, and every one of them is keeping score. Trust is the running total.

So we evaluate models over time, based on the kind of work the agent actually does, and we keep doing it as new ones ship. In the harness post I wrote about picking a daily driver for engineering. Same idea, higher stakes, because here the model's judgment shows up in every conversation the agent has. And like that brilliant colleague, you only really learn what a model is like to work with over time.

‍

Why Claude

Claude Opus keeps winning that job. It stays quiet in a channel where it has nothing to add. It knows the difference between context it can point to and context it should keep to itself. And when it speaks up unprompted, it's usually worth hearing.

That's the foundation we build on. Starting from a model with good instincts means we don't spend months coaxing basic judgment out of it, and we get to put our time where customers feel it: the identity, the tools, the guardrails, an agent handbook for each team, and the people who make every deployment work.

We use other models across the product where they fit, and every serious new release gets its turn at the job. Opus keeps earning it, which is the standard we'd hold any teammate to.

‍

The same kind of trust

We've worked closely with Anthropic while building Workforce. We meet with their team regularly, and a lot of those conversations are about exactly this: how a model behaves once it's actually on a team, and what we're seeing across real deployments. That relationship runs on the same things we ask of our agents: candor when something isn't working, and care with what's been shared. We also share a belief about AI at work. People should come out more capable, and stay in charge of the decisions that matter.

Between two teams or between an agent and the people it works with, trust gets built the same way, one small call at a time.