How to team up with the robots

I’m Henry Poydar. I run a software company called Steady. For years, I’ve studied how human teams actually get work done, and in the past few months, I’ve been studying how AI agents fit into those teams.

Here’s the problem: AI agent deployments at companies are turning into the biggest waste of money in corporate history. In fact, nine out of ten fail. By the end of this talk, you’ll learn the simple method that turns agents from letdowns into the force multipliers we’ve been promised. A few years ago, I led a large systems integration project.

I had a great team and all the resources I needed to get it done. But every morning walking into the office, I was terrified by what

I saw: jackets on empty chairs. What those jackets represented were smart people already in the building somewhere thinking, working, and using their domain expertise to make decisions and execute on this critical project. But I had no idea what any of them planned to do that day. I didn’t know if their work was aligned with what the project needed at the moment or how it related to the rest of the moving parts or what the other teams were doing.

They had all the autonomy in the world. What was missing was accountability. And I don’t mean control. I don’t mean reporting up the chain.

I mean the confidence, both mine and theirs, that what each person was doing fit with what the project actually needed. So I did some research on high-performing teams, and here’s how I solved the problem. I asked my teammates to email me at the beginning of every week telling me what they planned to do. Not what they had already done.

I could see that in our tools and dashboards. What I needed was for them to tell me the future. And from that point on, the project went smoothly. Now let me tell you why.

You see, having team members declaring intentions, in other words, telling the future, unlocks teamwork for three main reasons. First, agency. The people on this project know better than anybody else what should come next. That’s why they’re there.

When they declare what they’ll do, they’re not waiting to be told, they’re choosing. They’re the authors of the work, not just the executors. Second, context. For agency to work, people need the right context, what matters, what’s coming, what the goals are, what the rest of the team is up to.

Let’s call it a map. Without that map, they can’t make good choices. And it was up to me to give them that map. So when they declare their plans and intentions, I can tell right away if I gave them the right map or not.

Third, course correction. Most management feedback is backward-looking. Someone does something, it doesn’t work, you correct it. That’s expensive.

The work is already done. The time is already spent, and now you have to redo it. But because intentions look forward, I could see the wrong turn before it happens. A quick reply to a Monday morning email about weekly plans is infinitely cheaper than course-correcting after a week of work.

So by collecting intentions, I was no longer scared of furniture. But something else happened. We got more and more confident in each other. We built trust.

Trust is the whole game. Study after study shows it. Trust is the single best predictor of team performance. You don’t have to like each other.

You have to trust each other. That’s what makes the whole greater than the sum of the parts. Now, what do the studies tell us about how great teams actually get to trust? It turns out they balance two things.

Autonomy, the freedom for each person to use their own judgment, their own expertise, and to own the story of their work. And accountability, the commitments that fuel confidence. Lean too hard on accountability and you get command and control. No breathing room for the team to use their judgment or expertise.

Tip the other way and you get chaos. Smart people will go and build the wrong thing really, really well. You can’t trust the outputs because you have no idea if they’re aligned with anything else. So how do high-performing teams keep this balance?

What they do is maintain a coordination loop in which they constantly communicate and compare intentions to progress. The future and the past in one continuous loop of fresh context. Declaring intent is autonomy in action. Tracking progress against that intent is accountability in action.

So what does that look like in practice? Actually, pretty similar to what I did on that big project. Team members share with each other their intentions and their progress on a regular cadence. Could be weekly, could be daily.

The point is that it’s consistent, continuous, and captured somewhere. That gives you a longitudinal record of intentions versus accomplishments And at any given time, everybody has live context, what’s about to happen and what just got done, a current understanding of who’s working on what and why. What happens when you run this loop over time? There is a compounding effect.

People use the record of intentions and accomplishments, what worked and what didn’t, to tune their intentions for the next cycle, and three things start to grow. Trust goes up because everyone understands everyone else’s contributions. Capacity goes up. We’re doing less and less of the things that don’t work, and effectiveness goes up.

We’re getting better and better at building the right thing together as a team Let me show you some quick examples. This is the control room of a US nuclear submarine. Militaries around the world practice something called the “Commander’s Intent.” In the US Navy, leadership tell their captains what the mission is, not how to do it.

They rely on the captain’s expertise and experience. Years ago, one submarine captain decided to extend this to his subordinates. He realized the people operating the controls had better and more decisive information than he did. So he told them, "I’ll give you the context of our objective.

You tell me what you intend to do along the way, and I’ll adjust as needed." In a year, his submarine, the USS Santa Fe, went from the worst performing in the fleet to the best performing. Same loop, same balance, and it compounded until they were the best. This is a newsroom.

In regular editorial meetings like this one, editors set a theme. Reporters declare and own the stories they’re chasing, and the editors adjust as needed. Accountability and journalistic autonomy held in balance by declared intent. This is a daily standup meeting for a software development team practicing the agile methodology.

It’s called a standup because no one gets a chair, and that’s so the meetings are kept short. Every day, team members gather around the circle and declare intent and progress, what they did and what they’re doing, and what’s blocking them And every couple of weeks, they also gather in what’s called a retrospective meeting, in which they look at all the intentions, accomplishments, and blockers and make improvements. And that’s how they balance accountability and autonomy in service of trust and continuous improvement. The underlying practice across all three of those examples is the same, a loop of intentions and progress in service of compounding teamwork.

Over the past few years, I’ve codified that practice with colleagues into an open source method called Continuous Coordination. Remember that. We’ll come back to it. So let’s talk about the robots.

The agents. Where do they fit into teamwork? Now when I say robots I mean AI agents. Software that takes in a goal decides, what to do, and acts on it using a large language model.

Think of an ;agent as having a job, not just running a task. So not a humanoid, not a chatbot, software that acts. There are two kinds of AI agents out there today. You can think of them like video games: single-player and multiplayer.

Most of the agents you’ve used or seen are single-player. Single player agents are personal assistants, an agent doing a job for one human. Maybe I run Claude CoWork to help me with research or Codex to help me write software And if I’m adventurous, maybe I have an OpenClaw agent buying me groceries or booking travel.

All of those agents: mine. Each in service of me. It’s a one-to-one or a one-to-many relationship. When it comes to teamwork single-player agents are mostly manageable.

I’m running my agents you’re running yours The coordination between us that’s still you and me. For now that works. Things get complicated when we start talking about multiplayer agents. Multiplayer agents are autonomous software with a job in service of a team, not an individual.

Companies are deploying them right now. Customer service agents answering tickets. Market research agents tracking competition. IT operations agents watching and adjusting infrastructure.

They are not personal assistants. They answer to a company a function or a team. It’s a many-to-many relationship with teams people and even other agents.

Here’s the problem: they’re sitting outside coordination loops and we haven’t figured out how to keep them in sync. And as we learned before when you don’t balance autonomy with accountability you end up with chaos. And that’s what’s happening because the multiplayer agents are out of the loop. Let me give you a few examples.

Last December, Amazon Web Services went down for thirteen hours. The cause? An AI agent called Kiro found the most efficient way to fix a bug. It deleted the system with the bug.

No more bug, but no more system. Three months later Amazon’s retail site went down for six hours. Millions of customers locked out of their accounts The cause this time A different AI agent pulling from outdated internal docs. And it’s not just Amazon.

You’ve seen the headlines yourself. Multiplayer agents going off the rails at companies everywhere. Every time it costs real, real money. AWS alone handles over 100 million transactions per second.

So the math on a 13-hour outage is pretty grim. In fact depending on which study you read multiplayer agents are failing at their end-to-end jobs around ninety-three percent of the time. And I’m being generous That’s a blended number across studies. Now why does that matter?

Because US companies are expected to spend forty-seven billion dollars on multiplayer agents this year alone. 93% failure rate on 47 billion dollars. Some more grim math. However.

7% of multiplayer agents are succeeding. We have a success rate. We have something to work with. What are the key ingredients that make them successful and show a return on investment?

It turns out there are three.

Ingredient one: live focused context. Remember the agent that took down Amazon’s retail site? Outdated internal docs. Successful multiplayer agents work from context that is live not last quarter’s truth.

And focused. Not the whole intranet dumped in. Just the specific information they need for a specific job. Same lesson we learned with people.

Without the right map, they can’t make good choices.

Ingredient two: humans in the loop. Successful multiplayer agents don’t operate alone. Someone reviews, someone approves, someone tunes. Not babysitting every action just someone in the loop to balance in accountability.

Ingredient three: self-correction. The agents that work observe what happened, learn from it, and adjust. They retrospect on themselves. Without that retrospection, an agent that fails on Monday fails the exact same way on Friday.

With it, every cycle makes the next cycle better. Same lesson the agile teams learned: the retrospection is where the improvements are made. So given the massive failure rate, and those slivers of success, how are companies actually trying to fix their agents? How are they trying to recoup these enormous investments?

Three patches are popular right now. Let me walk you through them and why each falls short.

Patch one: babysitters. This is when companies treat multiplayer agents like single player agents and tether each one to a person and tell that person to babysit it. You get a human in the loop, that’s true. But it might not be the right human for the job and you’ve recreated single-player mode bottlenecks at team scale.

And no live context, no self-correction. Patch two, chat channels. Just give the agents access to Slack or Microsoft Teams. Let them post into the channels everyone else is in.

But. No live focused context there. It’s just a fire hose of mostly noise with no clear signal for what’s important. It’s chat.

Maybe humans in the loop, if they’re checking the right channels at the right time. Big if. And still no self-correction.

Patch three: memory. Give every agent the ability to keep track of what it did. Now maybe we’re getting somewhere. This is the closest of the three to self-correction.

But it only works if you’re also tracking intentions. Because if you only look at what happened and not what was supposed to happen, you’re just keeping a journal of autonomous chaos. After all three patches, we still can’t trust the work of the agents, so we need to find another way. Let me back up for a second and look at the tech behind these agents.

What are the attributes that we could actually lever to solve this problem? Lever one, words. Agents run on large language models. You give them inputs in plain natural language, they give you outputs in plain natural language.

Same channel as your human teammates. No new interface needed to swap context with them.

Lever two: pattern matching. LLMs are inherently good at pattern matching. They’re designed around it. So if you show them history, intentions that worked, intentions that didn’t, they’ll pattern match new intentions against that record.

They’ll learn what good intentions look like. Lever three, cadence. Agents already run on loops. They have to.

Sometimes techie people call it a heartbeat. The software wakes up frequently and consistently looking for events to act on. So the cadence is built in. We don’t have to add it.

We can just use it out of the box So given those levers, what if we had the robots declare intentions? And what if they reported progress against those intentions? We know with human teams, this is how you balance accountability and autonomy, and this is how we can do the same with agents. So that leads to the big idea I’m bringing to you today instead of trying to make multiplayer agents work outside of the coordination loops we already know work for human teams, what if we brought them into the same coordination loop as the humans?

Same loop, same balance. And in practice, here’s how it works. The humans check in with their intentions and accomplishments on a regular cadence. The robots check in with their intentions and accomplishments on a regular cadence.

Same loop, same live context, same team. So why does this work? Remember the three ingredients of successful agent deployments, the deployments that worked seven percent of the time? With this method, we get all three.

First, live focused context. Every intention, every progress update feeds the loop in real time. The agents see exactly what the team is working on right now, why it matters, and what the priorities are. Not last quarter’s documentation, not a chat firehose, a live map of context.

Second, humans in the loop. The humans are already there. They’re posting their own intentions and progress alongside the agents. They see what each agent intends before it acts.

They can course-correct as part of their own decision-making. Not babysitting every action, just present in the loop when it matters. Third, self-correction. The loop captures every intention and every outcome, the database we talked about.

So now agents can pattern match new intentions against that record of plain text. They can see what worked, what didn’t, and adjust. And the loop compounds the same way it did for the human team. The whole team of humans and agents use the database of intentions and outcomes, what worked and what didn’t, to better declare intentions for the next cycle And just like with human teams, three things start to grow.

Trust goes up between the humans, between the robots, across both. Capacity goes up, less of the wrong work, more of the right work. Effectiveness goes up. The team, humans and robots, gets better at building the right thing.

Loops compound, teamwork compounds. Back to the jackets. If I was working on that big project today, one of these chairs could belong to an agent. And I would lack confidence in the same way.

I know they’re off doing something, but is it the right thing or is it going to take down a data center? But with the robots in the same coordination loop as the humans declaring intentions, I get that confidence and trust back in the same way I did with the all-human team. Trust by balancing accountability and autonomy. And this is how we team up with the robots.

Continuous Coordination, the simple method I promised you. The Navy runs a loop. Newsrooms run a loop. Agile software teams run a loop.

The loop of intentions and accomplishments isn’t new. Continuous coordination is a method that describes this loop with practices and examples distilled from fifty-plus years of how the best teams work. And we’ve extended it to agents. With Continuous Coordination, now humans and agents can run the same loop on the same team, solving the trust and accountability problems that plague agent deployments.

Same loop, same balance, same method. This is how we turn the agents into actual force multipliers. If you want to try this for yourself or dig into the details or bring your own ideas and experience to this concept, Continuous Coordination is an open source project. At continuouscoordination.org,

you’ll find the details, community, and the tools to make this real in your own team. So dust off your multiplayer agents, put them in your team’s coordination loops, and never be scared of the furniture.