Don’t Manage Agents Like Employees

AI is turning everyone into a manager; people will spend far more time directing work than doing it themselves.

Today, AI mostly looks like virtual employees. Tag an agent in Slack, it updates Jira tickets, and comments on PRs in GitHub. It makes sense: that’s how you already interact with your coworkers, and now you’ll just have more AI direct reports to help you do your work. Most of the agents today follow this model, from Claude Tag to Cursor. 

But we think that’s not good enough. The scale, speed, and capabilities of agents are different enough from humans that interacting with them like regular coworkers will limit their potential.

We recently wrote about a startup with 10,000 engineers. When agents provide unlimited capacity to build, the constraint is human attention. We won’t be able to manage hundreds of direct reports with existing tools: running agent-native organizations will require agent-native collaboration tools.

Superhuman Workers Need Superhuman Review

A sophisticated AI user today might juggle 5 or 6 agents at once; tabbing between each session to keep them working. What happens when that number grows to 50-60? Agents are getting faster, and teams are setting up self-driving systems – from software factories, to marketing analyses, to investment researchers – that spawn dozens of agents without any human trigger. 

Agents will of course do some review themselves. But we still need people to set priorities, clarify intent, and be accountable for outcomes. With unlimited agent capacity, human attention will always be the limiter: we're going to be buried in outputs to review and decisions waiting for our judgment. 

Today’s collaboration platforms are designed for a human cadence. Engineering tickets move through development over the course of a sprint. Finish an hour-long meeting and you might have a handful of Slack threads to follow up on. 

Now imagine these at agent speed. While you were out for lunch, a ticket might have been created, specced, built, shipped, and marked complete before you get back to see it. Your Slack might have 200 agent messages waiting for you. The amount of work will be paralyzing. 

The paradigm of chronological message and ticket queues will need to be replaced by platforms that synthesize, prioritize, and direct our attention to the right things, at the right time, with context to make a quick decision.

A Communication Explosion

At any busy company, communication can already be a mess. Things get lost in DMs, channels get too noisy to follow, and decisions get buried in threads.

Agents will exponentially increase the scale and complexity of communication: 

• Agents write much more than humans; they work faster, don’t take breaks, and spell out lengthy internal narratives. Thinking traces and agent-to-agent conversations will be far too large to monitor. But they sometimes include key pieces of information that you might need to find. 
• Communication is a critical piece of context. Humans might know that an old Slack conversation was superseded by a new strategic imperative, but for an agent the full historical record can be interpreted as a standing decision. Conversely, a person might have a hunch that the answer to a question is in a specific thread, while an agent struggles to find it.
• Conversations will span multiple forums. A common workflow might see an agent as a member participating in an open channel. You might then want to pull context from part of that conversation into a private chat, direct the agent to do some work, then send that information to a different thread with another human coworker. 

Slack wasn't built for this. Its visibility model is coarse: everything in a channel is public to its members, and a DM is visible only to its participants. Conversations can either be in the main chat, a thread, or a DM; there’s no branching, threads-within-threads, and limited ability to move information between conversations. It works for a couple dozen humans, but for thousands of agents we’ll need new communication primitives to deal with the conversational explosion. 

The Limitations of Text

Anyone who has worked at a remote company knows it can be frustrating to communicate over text alone. How many Slack arguments were resolved with a quick call? 

People already collaborate in much richer ways than text. We get on video calls, sketch on a whiteboard, and walk through mocks together. But agents limited to text in tickets and messages are constrained to the lowest-bandwidth channel we have.

We’re not sure if agents will join the daily standup call (maybe they will!), but it’s clear they’re capable of communicating in much richer ways than text. No person would build an interactive prototype to answer a single question, but for an agent it takes minutes, so it's worth doing even if it gets thrown away.

The platforms we use to collaborate at work were natively built for text and images. Future tooling will need to be artifact-native: whether attaching comments directly to components in an working prototype, making multiplayer edits in living documents, sharing videos of product walkthroughs.

New Tools for New Companies

Today’s collaboration platforms are designed for human-shaped organizations: managers with a handful of direct reports, and work that’s done at human speed. 

Agents won’t just be drop-in replacements in the existing structure. Agent-native organizations will be much larger, have much broader management spans, and move much more quickly than was ever possible. 

Today, teams pushing the frontier find the limitations of our existing tooling, and build bespoke software around it. But we think there is the opportunity to build foundational new platforms that help the next wave of businesses manage an exploding agent workforce.

If you're building collaboration tools for teams of humans and agents, I'd love to hear from you: at@theoryvc.com.