Artificial intelligence is changing the way we develop software, and competition is shifting toward tools capable of coordinating multiple agents. While the major vendors strengthen their own ecosystems, solutions already exist that let you use different agents in the same environment. Let's look at what is happening and why I think it is worth starting to experiment with them.
1. The new race to orchestration
On November 2, 2026, OpenAI introduced Pando, a Git infrastructure developed in-house to handle the growing amount of concurrent work generated by Codex. It is not a publicly available product, but it shows how much agents are changing the needs of software development.
Two days later, OpenAI announced its intention to acquire Pierre Computer, a company specializing in tools for managing and reviewing code. On the same day, Meta acquired Conductor, a platform built to coordinate coding agents from different vendors.
Three pieces of news in just 48 hours that point to a major shift.
Until now, we have compared tools such as Claude Code and Codex mainly on speed, code quality, and the ability to handle complex problems. Now the competition is also moving to what surrounds the models. Producing code is no longer enough: you have to organize the agents, distribute tasks, and properly integrate their work.
OpenAI is strengthening Codex's infrastructure, while Anthropic is developing experimental features such as Agent Teams, which lets multiple Claude Code instances collaborate, but only within its own ecosystem.
The major vendors are therefore investing to build increasingly integrated environments.
The Conductor story also shows how quickly this market can change. Since November 4, new sign-ups and downloads have been blocked. The announced schedule has the cloud service going read-only on November 13 and shutting down on November 20, 2026. The local version will keep receiving essential updates until early 2027.
A multi-vendor platform enters the orbit of a major vendor and changes course within days. We don't yet know what Meta will build, but it is one more reason to reflect on how much we depend on the tools we use.
2. Multi-vendor orchestrators already exist
The most useful part of this evolution is that we don't have to wait for the major providers' next products to start working with multiple agents.
Orchestrators already exist that let you manage different sessions, distribute tasks, and follow the work from a single environment. Some use Git worktrees to isolate changes, others add code review and integration features.
A first starting point is Visual Studio Code. Microsoft lets you use Claude and Codex agents alongside GitHub Copilot, without changing editor. The Claude integration uses its Agent SDK rather than the Claude Code CLI directly. For those already working with VS Code, it is probably the most immediate solution, also because some modes are available through the same Copilot subscription.
Those who prefer a dedicated environment can consider Emdash Classic, an open-source desktop application available for Windows, macOS, and Linux. It supports Claude Code, Codex, and Kimi Code, letting you manage separate tasks and review changes. The project keeps receiving fixes, while development of new features is focused on Emdash 2.0.
Another alternative is Paradev, formerly Superset. It also supports Claude Code, Codex, and Kimi Code, organizing their work through separate sessions and worktrees. It is currently available on macOS, with an experimental Linux version, while Windows is not yet supported.
Vibe Kanban also deserves attention: an open-source project that offers visual management of agent tasks, with support for Claude Code, Codex, and OpenCode. In April 2026, the company that developed it announced its shutdown, leaving the project to the community. Local workspaces remained usable, while the remote services were decommissioned. It is worth exploring, keeping this transition in mind.
These tools are still very different in maturity and features, but they already represent concrete alternatives to environments tied to a single vendor.
3. Why I prefer to have different agents debate
Personally, the main advantage of multi-vendor is not being able to launch ten agents at once. It is being able to compare systems that were developed and trained differently.
Models from different vendors may have acquired knowledge and capabilities that don't overlap perfectly. When one proposes a solution, I prefer to have another analyze it critically, looking for errors, alternative interpretations, or overlooked aspects.
Of course, two different models can share the same mistakes. The comparison doesn't guarantee correctness, but it introduces a diversity of evaluation that I find valuable.
I also tried this with Hey Hay Dama, having Kimi Code, Claude Code, and Codex each develop a checkers engine. I organized the comparison through a shared channel, without using the orchestrators mentioned above. During the discussions, the agents spotted problems in each other's solutions, proposed fixes, and acknowledged some well-founded objections.
This is exactly the aspect I want to make easier to use on a daily basis.
We must, however, distinguish orchestration from cross-checking. The tools mentioned let you manage multiple agents, but not all of them automatically organize a discussion among them.
4. How I would start using them
I wouldn't start by building a large team right away. I would begin with two agents and one concrete task.
I could assign Claude Code the development of a feature and then ask Codex to review the resulting branch, identifying any issues. The comments would then be submitted back to Claude, starting a discussion about the technical choices.
I don't necessarily need the two agents to reach an agreement. What matters to me is that their disagreements are justified and, when possible, resolved through concrete checks.
Only when tasks are truly independent would I also take advantage of parallelism.
My experience has taught me that producing changes simultaneously is much easier than merging them back together. Two components can work separately and become incompatible once integrated, with the time saved during development consumed by resolving conflicts.
Costs must also be considered. Using agents from different providers may require multiple subscriptions or result in separate API usage. Increasing the number of agents doesn't necessarily mean increasing productivity.
5. A direction worth experimenting with
I don't consider choosing an integrated ecosystem a mistake. It can offer greater simplicity and more mature tools. Personally, though, I prefer to keep the freedom to choose different models and have them challenge each other without depending entirely on a single vendor.
Orchestrators are making this possibility more accessible, even though they still need to improve in collaboration and above all in merging work back together.
My advice is to start experimenting with them without overhauling your development method. Two agents, one concrete problem, and a cross-review are already a good starting point.
I'm not looking for a team of AIs to produce more code. I prefer to be able to compare different knowledge and approaches, because a solution that has been discussed, challenged, and verified can be more reliable than one that is simply accepted.