TL;DR:
- AI agents are joining teams fast, but the vast majority of pilots are failing – mostly because of coordination breakdowns with humans.
- Teams run on two coordination loops: ground-level (day-to-day sync) and big-picture (plans to progress). Teams were already struggling to close these loops before agents entered the picture.
- Agents that join teams need to participate in the same loops as their human teammates, or they fall out of sync, create more work, and wastefully burn tokens.
- Steady is extending its coordination system to include agents as teammates, backed by an open methodology called Continuous Coordination. Join the beta waitlist here.
AI agents are joining teams, and it’s kind of a mess.
McKinsey has 20,000 agents running alongside 40,000 human employees, targeting 1:1 parity by year-end. Salesforce closed 18,500 Agentforce deals across 124 countries. Cursor just announced autonomous triggering capabilities for its popular coding platform.
Last year, less than 5% of enterprise apps had embedded agents. Gartner says 40% will by the end of this year. Agents are answering phones, making decks, qualifying leads, processing invoices, writing code, scheduling interviews, and managing incidents. The average enterprise already runs 12 of them.
How do these agents stay in sync with teams of humans? Today, they mostly don’t.
The vast majority of AI agent pilots are failing, and 79% of those failures are directly tied to coordination breakdowns with the people they’re meant to serve. Teams added agents and got more work, not less. Incident management toil actually increased despite AI adoption.
Coordinating across people and teams is hard enough as it is. Now we’re adding agents as teammates, and it’s making the problem much, much worse.
Teams need two loops
Every team and company runs on two foundational coordination loops.
The big-picture loop connects plans to progress across teams and companies. Are we on track? What changed? What are the risks? Why does this work matter?
The ground-level loop keeps teammates in sync day-to-day with each other as they work on the big-picture stuff. What are you working on? What’s in your way? What do you intend to do next? This is the pulse of the team.
The coordination status quo for the loops (without Steady) is the “duct-tape stack”: meetings, emails, chat threads, dashboards, and manual toil stitched together into something incomplete, inconsistent, and inefficient. 55% of knowledge workers’ time goes to busywork. 25% of the workweek is spent just searching for information. 80% of the global workforce says they lack the time or energy to do their job.
The loops are barely closing with just humans in them, with teams piling up coordination debt and defaulting on it every day.
And now agents are here. They let people move faster, but faster also means falling out of sync faster. If duct-tape stack can’t keep people coordinated, it definitely can’t keep people and agents coordinated.
Single-player vs. multi-player
Most of the AI hype is around single-player agent mode. You open Claude, Cursor, or ChatGPT. It helps you write, research, code, think. The coordination endpoint is still you – you relay context to the team, you share the output, you close the loop. OpenClaw will send your emails, debug your code, and call restaurants on your behalf. ACowork organizes your files, tidies your inbox, and manages your calendar. These are genuinely useful. They’re also running in single-player mode. Your agents.
Even when the tools start to blur the line into teammates – Cursor’s new automations spin up background agents triggered by Slack messages, Linear issues, or merged PRs – the coordination model is still ad-hoc. Results get dumped into Slack or Notion. All the humans have to be watching and synthesize this into the team’s working context themselves.
So the harder coordination problem – the one we’re tackling – is multi-player agent mode. These agents aren’t helping one person. They’re acting as teammates. Multiple people interact with them. They do work that affects everyone. And critically, they need to participate in the same coordination loops as the humans.
Multi-player agents are here. Devin writes code for your engineering team. Alice prospects for your sales team. Rosie answers calls for your office. Olivia schedules interviews for your hiring team. Victoria processes invoices for your finance department. And then there’s all the custom agents built internally and by the likes of Glean or Salesforce. They serve whole teams or departments, not just one person.
But every one of them reinvents team coordination from scratch – CRM syncs, Slack notifications, tagging systems, confidence thresholds, approval chains. Each builds its own fragile mechanism for keeping humans in the loop. None of them participate in the team’s actual coordination loops, where real-time context is shared and compounded.
Do you know what your agents are doing right now? Do you know what they’re about to do next? Agents never check in, never surface blockers, and crucially, never express intent. What’s going to happen is the most important management signal of them all.
If we’re wasting 55% of human time in “work about work” already, how will that translate into wasted tokens for our agent teammates?
The agent rework problem
If you’re running multi-player agents today, you’re accumulating agent rework whether you see it or not. Every agent operating outside a coordination loop is a teammate who never comes to standup and never gets told when priorities change. It keeps working. It just keeps working on the wrong things.
The cost isn’t only tokens (though that’s real money). The bigger cost is the human time spent figuring out what the agents did, what needs to be undone, and what needs to be re-done. Not because the agents were built badly, but because nobody had a way to keep them in sync when context shifted.
Most companies are still wrapping their heads around agents as teammates. The ones who get ahead of agent rework will have a compounding advantage, because their agents and humans actually stay coordinated as things change. A well-coordinated team of people and agents will outperform a bigger team running agents that quietly drift off course every time someone changes a plan or tweaks context.
We’re building and testing a solution now
Steady is a coordination system for teams. Now we’re extending it to include agents as teammates.
We started with a conviction: teams need two coordination loops, and those loops should close automatically. That’s what Steady does for people today. Smart check-ins capture what you intend to do, what you accomplished, and what’s in your way. Goal story updates connect progress to the big picture. Agents distill it into personalized intelligence for every person on the team.
But here’s what we realized: multi-player agents need the same loops as people. Not separate agent-only dashboards. Not a new monitoring tool. Not a dump into a noisy chat channel. The same ground-level and big-picture loops your human teammates use.
So that’s what we’re building and testing now. In Steady, agents will check in just like people do. They’ll provide goal story updates. They’ll participate in both loops, alongside humans. Everyone gets visibility over what agents are doing, in the same place they get visibility into what people are doing – instant shared coordination memory and context.
We also open-sourced the underlying methodology. Continuous Coordination defines how people, teams, and agents coordinate through two async loops. The open schema makes it concrete – tool-agnostic, domain-agnostic, and designed for any combination of humans and agents.
This isn’t another SaaS inside a stack, it’s a system above the stack – the umbrella that keeps team actors coordinated automatically, whether they’re human or machine.
Get involved
Continuous Coordination is published here. The schema is on GitHub.
Steady is the system that implements it. We’re extending Steady to include agents as teammates in the same coordination loops as people, and we’re looking for teams who are already feeling agent rework pile up.
If that sounds familiar, join the beta waitlist. We’d love your input on what we’re building.
