From One Agent to Many

So working with an AI agent has been pretty revolutionary for me. It has allowed me to automate and run jobs and workflows that would have normally occupied most of my time, now in a much simpler manner, independent of my time invested. This isn't news; I think everybody sees this experience shared by many online. What I'm seeing more and more now is the ease or the accessible ability to coordinate many agents that then obviously multiplies your effectiveness. This coordination of agents is happening in the spaces of software development, of personal agents, and even in a more abstract way within organizations.

I don't mean to add to the pressure that is very palpable for many of us, where there is a sense that you are being left behind. I do want to encourage everybody, with a strong sense of urgency, to start engaging with these tools, because it takes some time to truly understand the shift in possibilities and to understand the realm of productive environments that are now available with this technology.

Now, agent coordination and orchestration is nothing that's necessarily new. There are examples of this going back a bit, but what is new is that I am seeing new projects come online and become available that make it really easy to implement these sorts of orchestrated environments for running multiple agents. So I have several examples of this and then one tangential example of how an organization has coordinated agents and people using agents at scale and thoroughly throughout the organization. I thought that this was an important use case to include here because the findings and results are pretty significant. This is another example of just how quickly things are evolving.

Coding Agents

The first few examples of agent orchestration are related to coding agents. The projects that I have found interesting that I have seen scroll across my various feeds include Superset, Cline Kanban, and mngr.

Superset allows you to have multiple terminal windows running with coding agents in a well-organized UI and additionally helps you easily create additional copies of your software development projects via automatic worktree creation. This allows you to run multiple tasks on a single project within a well-organized UI. Before this, I had multiple windows for my terminals, but this helps organize everything visually and helps you track, through visual cues, the status of each of your agents when they are running and when they have completed their tasks. This particular project has been super helpful to me.

Next is Cline Kanban. I have not used this project much yet. I keep forgetting to use it more extensively since I saw this late last week. But simply put, it also helps coordinate agents and tasks by breaking a project down into individual tasks. Directly from the product page, "Create, manage, and chain agent tasks within a single view."

Finally for coding agents today, let us look at mngr. Now I am not sure if I will make much use of this last tool because the solution itself is pretty advanced. But I'm including it here because it is impressive and pushes the boundary for how software development can be conducted. The MNGR solution allows you to basically spin up agents not just in your local environment but anywhere that you can connect to via SSH, a protocol for basically connecting via terminal to different computers. Now, this is how certain features in Claude desktop and Claude code work, but it happens automatically in the background with MNGR. You have control over how to run agents and basically coordinate a fleet of them running on different machines, which is pretty damn cool, but also maybe beyond my needs, or maybe just beyond my imagination at the moment.

Personal Agents

Within the last couple of months, there's been an explosion of what I would call autonomous personal agents that were popularized by Open Claw. This past week, I saw this new project called Paperclip, which is basically an orchestration tool for running several of these personal agents within one "control plane," as their website defines it. Maybe using the term "personal agents" is a bit misleading, because the intention behind Paperclip is for you to easily establish a company with AI agents operating in different roles for that company. You set high-level goals, metrics, and outcomes. The CEO agent figures out hiring, delegation, and execution. This is a deliberate shift from the "pair programmer" model most people use with Claude Code.

I've installed Paperclip but haven't used it yet, so I can't vouch for how effective and well-functioning the coordination of all these agents filling different roles is. But the demos I have seen look very impressive.

claw-code is up next. Now technically, this is related to coding agents or the coordination of coding agents, but this is very different from the coding agents that I mentioned previously because this demonstrates work that is completely autonomous. The implications of this are more closely aligned with general agent behavior rather than specifically that for software development. So if you did not hear, the source code for Claude Code was leaked last week, and within a day, or even shorter than that, within a night, someone had rewritten the implementation in a different programming language. That's what we're discussing here, the rewrite of Claude Code as claw-code, autonomously in a sense. I would encourage you to read the post from the developer who did this. There are some very interesting insights that he has, and the architecture, the setup and mechanisms around how he coordinated the development of this copy of one of the most popular coding agents on the market, which he did through Discord only, a chat app, is just jaw dropping.

Here is one of his insights:

But if you are staring at the generated Python files, you are looking at the wrong layer. The code is a byproduct. The Rust port that followed is also a byproduct. The thing worth studying in the claw-code repository is the system that produced all of it. claw-code was always a showcase. The point was never the Python files or the Rust crates. The point was the clawhip-based agent coordination system that built them while the developer was asleep.

Institutionalized Agents

In American English, when you say someone is institutionalized, it means that they've been put into a mental health facility. I thought that the sub-title was a good fit here because what Trail of Bits has done will seem crazy to anyone that has struggled to effect any sort of change at any institution that is not small.

Trail of Bits is a cybersecurity consultancy with about 140 people. Over the course of roughly one year, they went from 5% of staff actually using AI to having AI-augmented auditors finding 200 bugs per week, up from 15. Twenty percent of all reported bugs are now initially discovered by AI. Their revenue per sales rep hit $8 million against an industry benchmark of $2-4 million. These are not small improvements. This is a structural shift in what a 140-person company can output.

What makes their story worth including here, alongside individual tools like Superset and Paperclip, is that they solved the coordination problem at the organizational level. It wasn't just about giving people access to AI. Trail of Bits built what they call a six-part "operating system" instead.

The short version is that they standardized everyone on the same tools, wrote an internal handbook explaining not just the rules but the reasoning behind them, created a maturity matrix so people could see where they stood and where they could grow, ran focused adoption sprints that forced people into autonomous mode, captured every engagement's learnings as reusable artifacts, and then made autonomy safe through sandboxing and hardened defaults pushed to machines via device management software. By the end, they had accumulated 94 plugins, 201 skills, 84 specialized agents, and over 400 reference files. Every project made the next one faster.

The reframing of identity resonated with me particularly well, having already experienced resistance myself. Their security auditors are experts, and experts resist AI harder than anyone because their skill is core to who they are. Trail of Bits didn't fight that, but instead side shuffled past it. They reframed AI proficiency from "I use AI" to "I'm the one who makes the AI dangerous." That's an identity upgrade, not a threat. It worked because it aligned with how these people already saw themselves. Beautifully executed, Trail of Bits.

Where This Leaves Us

This is the direction things are heading. The examples earlier in this post show how individuals can coordinate multiple agents. Trail of Bits shows what happens when an entire organization does it systematically. The gap between companies that figure this out and companies that don't is going to get very wide, very fast. I'm treating this less as something to watch and more as something to build toward, both in my own workflows and in how I think about bringing this to my team at work as something that I hope to template and then roll out to the entire organization.