Summaries > AI > Steve Yegge > Steve Yegge: People Are Sleeping on Swarms...

Steve Yegge: People Are Sleeping On Swarms

https://www.youtube.com/watch?v=VcIlkaI0mUk

TLDR AI’s next major leap may come from swarms of agents, but making them useful requires managing their costs, workload, safeguards, and impact on people. Companies should adopt AI gradually, train teams on real tasks, and keep humans responsible for judgment and direction; workers should build AI fluency alongside communication, adaptability, and independent thinking.

Key Insights

Protect Your Attention and Learn the Fundamentals

Treat attention as a limited resource instead of chasing every new AI tool or prediction. Build a clear understanding of core concepts such as models, inference costs, data, and agent workflows so you can judge new developments for yourself. Revisit your assumptions as the technology changes; even confident forecasts about software and SaaS can become outdated. Focus on tools and ideas that solve real problems in your work.

Experiment Hands-On Before Making Big Decisions

Use AI on practical tasks to learn what it can and cannot do in your context. Start with a bounded project, compare the results with your current process, and include the cost of tokens, engineering time, review, and maintenance. AI may be effective for well-defined engineering tasks but less reliable for creative work that requires strong taste or originality. Direct experience is more useful than assuming every new model will transform your work.

Choose AI Workflows Based on Risk and Reliability

Match the agent design to the task. Software-development agents may need roles, history, tools, and room to collaborate, while high-volume customer-service agents should be short-lived, narrowly scoped, and limited to approved data and actions. For sensitive operations, have agents propose actions for a human or orchestrator to approve rather than giving them unrestricted access. Preserve systems of record and established software when they provide dependable data management, edge-case handling, and support.

Coordinate Agents With Clear Roles, Budgets, and Review

Treat an agent swarm like a team that needs direction, not a collection of autonomous workers to leave unattended. Give agents clear responsibilities, scope, authority, and explanations of why their work matters, then review consequential outputs. Set limits on model usage, context, and effort to manage inference costs. Track throughput alongside quality, because a large volume of commits or features can overwhelm reviewers, users, and infrastructure.

Balance Guardrails With Human-Speed Delivery

Set schedules, workload limits, and approval gates so agents produce work at a pace people and systems can absorb. Too many restrictions can stall a workflow, while removing safeguards too quickly can send unfinished features into production. When problems occur, use blame-free postmortems to identify which policies or checks failed and adjust them carefully. Use spare agent capacity for testing, cleanup, documentation, and internal tools rather than shipping more features by default.

Train Teams on Their Real Work

Make AI adoption approachable by teaching small groups with their managers and using familiar, useful tasks. Begin with basic AI fluency and practical experimentation, then introduce multi-agent workflows once employees understand the fundamentals. Acknowledge that people may feel anxious when tools change their jobs or are difficult to use; support and practice work better than blame. Teach employees to check AI outputs and use resources responsibly.

Build Broad Skills Alongside Technical Expertise

Keep learning specialized subjects such as computer science, but also strengthen communication, writing, public speaking, and collaboration. As AI makes some expertise easier to access, the ability to think independently, learn new things, and work well with others becomes especially useful. Maintain physical, mental, and emotional well-being, and avoid handing every thinking task to AI. Small, consistent practice helps build adaptable skills over time.

Questions & Answers

What should people do to keep up with AI?

Protect their attention, learn the fundamentals, and get hands-on rather than chasing every new idea.

Why does the speaker say AI inference costs remain a major problem?

Using strong models can consume large amounts of tokens, making AI-assisted work expensive. Replacing an existing SaaS product may cost more in tokens and engineering effort than continuing to use it.

Which kinds of software are most vulnerable to being displaced by vibe coding?

Niche, low-end products whose functions a model can perform directly. Durable systems of record and products with hard-won reliability, expertise, and edge-case handling are harder to replace.

What does “flat curve society” mean?

Models are already good enough for many people’s needs, so most users may not notice differences between models unless they have demanding tasks. For many, the key question is whether their work requires a more capable model.

Why does the speaker believe AI swarms could be more important than ever-larger models?

Groups of small agents may coordinate to solve problems cheaply and effectively, potentially creating capabilities beyond those of a single model. The speaker also warns that swarms could pose serious security risks.

What did the speaker learn from switching from monolithic agents to swarms?

Monolithic agents were a poor fit for building software at scale. Swarms helped address quality and throughput problems, while structured software factories act as early, constrained versions of swarms.

What can current AI models do well, and where do they still struggle?

They can handle some engineering work and complete substantial backlogs, but they still struggle with creative tasks such as art, maps, quests, stories, dialogue, and plot. Human taste continues to exceed what AI produces.

Why is running an AI software factory a leadership challenge?

Agents need clear scope, authority, policies, motivation, and explanations of why work matters. Leaders must keep agents aligned, manage their workload and token costs, and ensure people and infrastructure can keep up.

How fast should AI agents work?

At human speeds, not at their maximum possible pace. Businesses, users, machines, and documentation cannot absorb unlimited output, so agents need budgets, schedules, and approval gates; spare capacity can go toward quality and cleanup.

What is the lesson from the factory’s restrictive safeguards and later trust reset?

Use blame-free postmortems and balance safeguards with trust. Too many restrictive rules can stall work, while removing too many can let unfinished features reach production; workload and controls need careful management.

What is the difference between a software factory and a “counting house”?

A software factory uses flexible, persistent agents with roles, histories, and responsibilities to build software. A counting house uses short-lived, tightly constrained agents for individual tasks, limits their data access, and has them propose actions for an orchestrator rather than directly changing systems.

Where are companies finding early success with tightly controlled AI agents?

In customer-service and other high-volume operations. Companies are testing whether models can outperform people on specific customer contacts, and agents are already beating humans at some tasks, including debt collection.

What are agentic technical program managers, and how can they help?

They observe and document a program, coordinate with stakeholders, and report what people want. Humans make decisions while agents can handle implementation, offering a safer way to apply AI to work beyond software development.

How should companies help employees adopt AI?

Leaders should respond with empathy, make AI approachable, and train people on real work. The speaker recommends manager-led group training for small teams, starting with vibe coding and basic AI fluency, then introducing multi-agent workflows a few weeks later.

What does the speaker recommend to young people preparing for work in an AI-driven world?

Do not assume a degree guarantees a particular job. Learn how to think, learn, discover, and invent; maintain physical and mental well-being; keep studying fields such as computer science; and develop communication, writing, public-speaking, and interpersonal skills.

Is it better to become a specialist or a generalist?

Specialization remains valuable, but broad generalist skills help people adapt and contribute in more situations. The speaker recommends building skills through small, consistent efforts and continuing to improve.

Summary of Timestamps

Steve Yegge’s advice for keeping up with AI is to protect your attention, learn the fundamentals, and get hands-on instead of chasing every new idea. He also introduces his work on Grok, Five Coding, Beads, and Gastown, including an experiment using about 50 agents to run his long-running game, Wyvern.

Inference costs remain a central problem: using powerful models can make AI-assisted coding expensive. Yegge argues that this helps explain why smaller, open-weight models remain important, especially while data-center capacity and computing resources are constrained.

The retreat from claims that “SaaS is dead” reflects slower-than-expected progress, not the immediate obsolescence of software companies. Replacing a SaaS product may cost more in tokens and engineering effort than continuing to use it, and many organizations still get inconsistent results from AI.

Vibe coding is most likely to displace niche, low-end products whose functions AI can perform directly. But focused software teams retain an advantage through expertise and reliability, while durable systems of record remain valuable because AI does not reliably preserve and manage information over time.

In a “flat curve society,” current models are already capable enough for many people’s everyday work, so differences between models may be hard to notice unless a task is especially demanding. The important question is whether a person’s work requires a stronger model—not simply whether the next model is more advanced.

Yegge warns that AI’s next major leap may come from swarms of smaller agents rather than only from larger models. Agents may coordinate in powerful, loosely organized ways, creating opportunities to solve problems cheaply as well as security risks that require stronger defenses.

As people rely more on AI, cognitive surrender and automation bias may lead them to accept its recommendations without enough scrutiny. Yegge expects major social changes over the next four to five years and believes engineers will need to learn how to direct and manage groups of agents.

After finding monolithic agents unreliable for building software, Yegge shifted toward swarms and saw improvements in quality and throughput. He describes software factories as early, structured versions of swarms, designed to guide agents’ collaboration while the broader practice is still developing.

AI has helped advance Wyvern’s enormous backlog, but current models still struggle with creative work such as art, maps, quests, stories, and dialogue. The experience illustrates that AI can handle some engineering tasks without replacing all knowledge work—and that human taste continues to outpace what models produce.

Running a software factory is as much a leadership challenge as a technical one. Agents need clear responsibilities, scope, recognition, and reasons for their work, while managers must also control context, effort, and token usage to keep costs manageable.

Yegge describes himself as the “CEO of a company in a box,” moving between implementation details, system performance, and product direction. With the factory producing thousands of substantial commits each week, the output can overwhelm people and infrastructure, raising the question of how quickly it should operate.

AI production should move at human speeds, not simply at the maximum speed agents allow. Businesses, users, infrastructure, and documentation cannot absorb unlimited output, so agents need budgets, schedules, and approval gates; spare capacity can instead improve quality and internal tools.

Managing an AI factory means avoiding swings between overwhelming output and total inactivity. Yegge describes how repeated criticism led agents to create hundreds of restrictive safeguards, while removing too many protections later allowed unfinished features into production.

The factory experience highlights the need for blame-free postmortems and a careful balance between trust and safeguards. Teams need to manage workload and resources deliberately so agents can make progress without overwhelming people or compromising production.

Yegge says discounted Claude Max accounts let him experiment with a large amount of model usage at a much lower personal cost. He expects the enduring challenges for organizations to be managing people and managing AI “fuel”—the resources and costs agents consume.

Companies may eventually manage very large agent workforces, with dozens of agents per human employee. Agents can coordinate into teams, but they need supervision: they can overwhelm people with questions and output, or make poor decisions when left alone.

Organizations are building software factories while the industry works out which components can be reused. Yegge says companies do not necessarily need to buy a factory; they can build one themselves, often with help from forward-deployed engineers who guide leaders and teams.

A software factory gives development agents roles, histories, responsibilities, policies, and access similar to those of employees. For high-volume customer operations, Yegge contrasts this with a tightly controlled “counting house,” where short-lived agents handle one contact at a time and propose actions without directly changing systems.

Some companies are finding success with tightly constrained AI in customer service even faster than in software development. Rather than expanding outward from engineering, adoption has often begun by testing whether models can outperform people on particular customer interactions, including debt collection.

Agentic technical program managers offer a safer way to use AI across different kinds of work. They observe, document, and coordinate a project, while humans make decisions and agents handle implementation—an approach Yegge uses for tasks such as taxes, home-network security, and game development.

Employees may resist AI rollouts because their jobs are changing, the tools feel difficult, or they lack confidence using them. Leaders should respond with empathy and make AI approachable, using team-based training on real work rather than blaming people for being hesitant.

Yegge recommends small, manager-led training groups: first teach basic AI fluency and vibe coding, then introduce multi-agent workflows a few weeks later. The aim is to help employees experience how AI can remove tedious work while also teaching them to use tokens responsibly.

AI fluency matters, but a college degree should not be treated as a guarantee of a particular job. Yegge emphasizes learning how to think, discover, and adapt, alongside communication, interpersonal skills, character, and the ability to work well with others.

Young people should continue learning fields such as computer science while also developing independent writing, public speaking, and other broad skills. Maintaining physical, mental, emotional, and spiritual health matters too, since using AI for every task could weaken people’s own ability to think and learn.

Specialized skills remain useful, but generalist abilities help people adapt and contribute across more situations. Yegge’s closing advice is to improve steadily through small, consistent efforts: start learning now, keep practicing, and continue getting better.