Agent Orchestration: From One Person to a Team
One person can now hand work to several AIs at once. The next challenge is helping people who have each built their own AI team to work together.
If your team uses AI for real work, you may know this situation. A teammate sends over a long document packed with requirements and AI-generated output. For all its length, it leaves out why decisions were made and what happened along the way. It's full of unexplained terms, too. So you ask your teammate to walk you through it.
A few years ago, AI usually meant one tool giving you one answer at a time. Now one person might have one agent writing, another fixing code and a third looking things up, all at the same time. That person can clearly get more done. Whether the whole team is that much more efficient is another question.
This post looks at how AI orchestration has evolved, why the remaining challenges lie between people and what Riido does about it. It ends with a few practical steps any team can try, whatever tools they use.
1. How orchestration got here
Orchestration means dividing work among AIs and bringing their output together into a finished result. An AI that takes on a task and handles it independently is called an agent. Over the past few years, this way of working has gone through roughly three stages.
First, one AI worked on its own (2023). You gave it a goal, and it chose and carried out the steps. But if it ran without a person's approval, it could get stuck in loops, repeating the same work.
Next, several AIs took on different roles (2023–2024). People began assigning planning, development and review to different AIs. When researchers at UC Berkeley analyzed the execution logs, they found plenty of failures, such as agents repeating steps or not knowing when to stop.
Now, one person can run several AIs at once (2025–). Claude Code and Codex made it common to keep multiple AIs working in parallel. In February 2026, Claude Code released an experimental feature that lets several AIs share a task list and message one another.
But splitting work among several AIs doesn't automatically lead to better results. Google Research compared 180 AI setups. For tasks that break neatly into separate branches, such as financial analysis, one AI coordinating several others did 80.9% better than a single AI working alone. But for work that needs to happen in sequence, every multi-AI setup they tested performed 39–70% worse than a single AI.
An Anthropic experiment points to the same lesson. Researchers asked 16 Claude instances to build a C compiler, a program that turns human-written code into something a computer can run. When they tried to compile a program as large as the Linux kernel, all 16 agents got stuck on the same bug and kept overwriting one another's fixes. Only after they broke the task down again file by file, using a known-good compiler as a reference, could the agents split up the work.
In both studies, how the work was divided mattered more than how many AIs were involved.
2. One person now leads an AI team
For people who use AI well, the workday is starting to look a lot like a team lead's. They break work into pieces, hand it to several AIs, bring the results together and check them against a standard. To build a feature, for example, one AI might update the screens, another change the server code and a third write the tests. The person then combines the work and checks that the feature does what it should.
A team lead's job has four parts. They decide who does what, track progress, carry results from one piece of work into the next and check that the finished work is ready to use. The more AIs there are, the more time all of this takes. But when one person is running the AI team, the relevant information and context live in one person's head and on one computer, so things rarely fall out of sync.
Riido's product team has five people, one designer and four developers. As we shared in an earlier post about rebuilding Riido in nine weeks, we work with an average of 30 AIs.
With five people and 30 AIs, there are 595 possible relationships. Not every pair needs to talk, of course. What matters is how information actually moves. Within your own AI team, you direct the traffic. To reach a teammate's AI, information has to pass through you and your teammate. And there are just 10 connections between the five people.
Five people and 30 AIs can form 595 possible relationships. But one person's AI can reach another person's AI only through the 10 connections between people.
Some work/project management tools now let you assign an AI to a task. Still, most AIs work for one person, and the tools that coordinate several AIs keep everything within that person's reach, such as their computer or chat session. Claude Code's experimental feature, for example, stores its task list and team settings on the computer of whoever started it.
In a 2026 talk, Maggie Appleton, a researcher at GitHub Next, called all of today's AI coding tools "single-player interfaces." They're built like games you play alone. So no matter how big one person's AI team gets, the point where it meets someone else's team is still a person.
3. The next challenge: collaboration between AI team leads
When people who use AI work together, each person needs to pick up work from a teammate's AIs as well as their own. That's where four kinds of problems tend to show up.
Split: Who's doing what? When two people's AIs work on the same thing at once, their efforts can collide. In the compiler experiment, all 16 AIs piled onto one bug. In day-to-day development, the same kind of overlap means doing the same work twice. A 2026 study examined more than 33,000 pull requests that AI agents opened on GitHub. A pull request is a proposed code change waiting to be merged into a project. The researchers looked closely at 600 that were never merged. In 23% of them, the change duplicated one that had already been submitted.
See: What's happening right now? You can watch your own AI on your own screen. To find out what a teammate's AI is doing, you usually have to ask. When progress isn't visible to everyone, people may repeat work or wait for something that's already done. A marketer's AI might draft a post about a feature while a developer's AI is still changing that feature. One of them will have to redo their work.
Hand off: Can the reasons travel with the result? When your AI needs to build on something a teammate's AI made, the work passes through both of you. The result arrives, but why it was made that way and what was tried and rejected often get lost along the way. For example, a teammate's AI might leave an unreleased feature out of a newsletter draft on purpose. If only your teammate knows why, your AI may treat the omission as a mistake and put the feature back in.
Review: Who checks the result? AI can produce work quickly, but a person still needs to decide whether it's ready to use. In the same study, 38% of the unmerged pull requests were closed without a person properly reviewing them. That's far more than the 17% that failed automated checks. Without a clear decision about who reviews the work and when, the more work AI produces, the more review work piles up.
When you work alone, you handle the dividing, tracking, handoffs and checks yourself. Add even one more person, and those jobs spill into conversations, chats and meetings.
As the number of agents grows, the people connecting them become the bottleneck.
4. What Riido does
In Riido, each person can connect the AI they already use to tasks the whole team can see. An AI registered as a team member is called an agent. The requests it receives, its results and the replies that follow stay in the task record instead of living only in someone's memory.
With AI Agent, launched on October 22, and the November 5 update, you can do five things today. As an example, take a marketer and a developer working together on the November newsletter.
Register the AI you already use as a teammate. Connect Claude Code or Codex on your computer to the Riido desktop app, and it joins as an agent with its existing memory and settings. The AI still runs on your computer. In our example, the marketer registers Claude Code to write the draft and Codex to review it. The developer registers their own Codex. Each agent's Settings tab lets you choose its name, model and instructions, so you can set expectations up front. Changes from GitHub, Figma and Slack can also be linked to tasks.
Assign work as you would to a person. Choose an agent as the task's assignee, or @mention it in a comment. Before it starts, the agent reads the task description and comments, along with the parent project and milestone, subtasks, linked tasks and meeting notes. When the marketer assigns the "November newsletter draft" task, the agent also reads the project and milestone, the linked "Finalize release notes" task and notes from the linked weekly meeting. Each reply includes a "Sources read for this answer" list, so people can see what the agent read and used.
Let agents pass work to one another. When another role is needed, an agent can find a suitable agent and ask it directly. The second agent joins the same conversation with the earlier context intact, and everything stays in the task comments. In this example, the marketer's Claude Code writes the draft, then asks the marketer's review Codex to check it. The reviewing agent leaves suggested fixes as comments. The work moves forward without someone copying results back and forth.
The first part shows the November 5 update. A development agent reads the project, milestone, task and meeting notes, completes its work, then asks a review agent to take a look. The second part shows five people and 30 agents working together. Agents pass work directly to agents on their own team and on other people's teams. People see the results, discuss them and send work back to agents.
Keep follow-up requests in the same conversation. Reply under an agent's result, and it handles the request in that same conversation, still remembering what was discussed and how the work went. Even "Make the second paragraph shorter" works without re-explaining the task. When the developer opens the task later, they can see the result and the back-and-forth edits, and find answers there before asking the marketer.
Track progress in one place, and leave the final check to people. Assign work to an agent and a progress indicator appears. The agent detail page shows its active tasks, queue position and run history. Use the Assignee filter to find tasks an agent has worked on. The task's Activity tab keeps the agent's run history and explains failed runs. Agents can exchange feedback within a task, but the assignee decides whether to mark it Done.
Together, these features address the four challenges from the previous section. Task assignments show who's doing what. Progress indicators and agent views show what's happening. Task conversations preserve the context behind the work. And the assignee remains responsible for the final review.
In the newsletter example, the draft, the sources the agent read, the review comments and the marketer's edit requests all stay on the "November newsletter draft" task. Whoever handles next month's newsletter can open it and see what was decided last time and why. Link that task to the next one, and the new agent can read it before getting started.
5. What your team can try first
If several people on your team work with AI, here are a few things to try first, whatever tools you use.
- Record delegated work somewhere everyone can see. Note what you handed off and whose AI is doing it.
- Assign each piece of work only once. Use one shared place to assign work, and check whether someone is already on it.
- Don't split work that needs to happen in order. If an earlier decision shapes what comes next, have one AI handle the work from start to finish. In the Google study, dividing sequential work among several AIs led to worse results.
- Keep the reasons next to the result. Write down what was chosen and why, and what was tried and set aside. This helps the next person or AI avoid repeating the same trial and error.
- Put decisions somewhere AI can read them. If a decision from a meeting lives only in someone's notebook, the AI won't know about it. Keep decisions and standards in a task, document or other place AI can read.
- Set the definition of done and name the reviewer up front. Write down what counts as done and who will check it. Criteria written on the task also become instructions the AI reads.
It took only a few years to go from one person asking one AI for an answer to one person leading a team of AIs. AI can do much more on its own now. The next challenge is helping people who each have an AI team pick up one another's work. We think a good first step is to have people and AI work on the same tasks.
To connect the Claude Code or Codex you already use as a teammate, start with the Riido desktop app.