AI Engineer Summit 2025
Reverse Conway's law and GenAI: How agents will take over the organisation
Loading slides…
Read the talk
Reverse Conway's law and GenAI: How agents will take over the organisation
As agents move from personal copilots to organizational peers, the design problem shifts from where AI belongs to who specifies, reviews, governs, and benefits from its work.
From a talk by Patrick Debois
Where does AI belong?
Should generative AI belong to the data team, the application team, or the platform team? That is the opening organizational problem in Patrick Debois’s talk. The enthusiasm around ChatGPT has encouraged people across the industry to become AI specialists, but adding AI to existing work does not necessarily change what an organization can accomplish. Choosing an owner is already a decision about how the organization should change.
Debois previously addressed platform teams as a way to introduce and scale AI. Here, the starting point is Conway’s Law, developed in How Do Committees Invent?: the communication structure of an organization constrains the systems it designs. Small, distributed teams tend toward modular services because their communication boundaries influence the boundaries of their software. The mechanism is communication, which need not follow the formal reporting hierarchy.
Debois approaches this question through his background in automation and DevOps and his focus on engineering with generative AI. He offers the slides through LinkedIn in exchange for feedback and describes his work as a curator and contributor to the AI Native Dev community. The question he now turns toward is the reverse direction of influence: how does AI change the way people organize themselves?
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From typing code to specifying intent
The first change happens at an individual’s desk. A copilot becomes a sidekick to which someone can pass knowledge, then a contributor producing more of the code that enters the organization. That progression raises an uncomfortable question: if the system increasingly performs the work, what remains for the person? A company culture that supports learning gives people room to rethink their jobs rather than face that uncertainty alone.
For developers, the shift is from typing toward intent. Generating the next piece of text gives way to planning what should exist. Product requirements and specifications become inputs from which a system can generate code. The important work moves toward expressing the desired behavior clearly enough that another system can implement it.
Review does not disappear with typing. Someone still has to distinguish a good implementation from a bad one. The developer’s role consequently moves from creating each detail toward managing the work: less attention to how every line is produced, more attention to why the system should behave a particular way.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A job is a bundle of tasks
A job contains several tasks, and technology does not affect all of them equally. It can complement one task while substituting for another. Unbundling the job makes the possible changes easier to see.
| Change | Where the person’s value goes |
|---|---|
| Complementary assistance | More output, potentially commanding a premium |
| Value migration | From coding toward management or specification |
| Commoditization | Toward another task that remains valuable |
| Substitution | Potentially toward another job |
These are different possible outcomes of introducing technology. Increasing output in the current role is only one of them; the task that previously distinguished someone may instead cease to command much value.
One new task is teaching AI to perform the old task. Debois points to Outlier, where people are paid to contribute expertise to AI training, including coding expertise. His point is not that this is a bad job, but that the work can move from writing software toward training a system to write it.
The personal response can move through its own sequence: confidence that the technology will not affect the job, fear that it will replace the person, and eventually the discovery of things that person could not previously do. That emotional movement accompanies the redistribution of tasks; it is not resolved simply by giving everyone a copilot.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Automation still needs failure expertise
Two pricing bots on Amazon react to one another and drive the price upward. Debois uses this example to show how automated systems can create a failure through their interaction. Someone must understand what happened and clean up when the machines cannot resolve it. That is a plausible intermediate role for humans, though it does not guarantee that every automated task will retain a human operator.
Experience with automation makes designing for failure part of the job. Prevention, recovery, and continuing training remain necessary even if fewer people operate the system. This creates a paradox: the more routine work automation removes, the less often people may exercise the expertise they still need when something breaks. Debois raises the possibility that maintaining readiness could offset the gains, even producing a net-zero effect in some circumstances. He presents that as a possibility, not a measured result.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The assistant becomes a shared teammate
The next organizational step is an assistant that serves the whole team rather than one person. Human teams assemble overlapping expertise through full-stack engineers and specialists from different domains. Even a multidisciplinary team has a finite pool of knowledge; Debois contrasts that with the much broader range of subjects an AI can address. People can draw on that knowledge, but they still have to verify the answers.
If AI supplies some of the overlap previously provided by multiple people, a smaller team might cover the same domains. Its members would still need the judgment to recognize good work, so reduced headcount would also change the required skills. Debois reports seeing teams grow as well as shrink; the direction of the size change is unresolved, even as the composition changes.
Alongside the application and infrastructure stacks, the organization may acquire a teammate stack: agents with particular strengths in sales, data, and other domains. Selecting these teammates becomes another way of assembling the expertise a team needs.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Faster execution changes the human loop
Faster agents do not remove the need for decisions. One person might oversee several assistants, stepping in as a specialist when something fails or requires judgment. Execution can spread across more workers while decisions still converge on the human.
Debois imagines one- or two-week development cycles shrinking to two days. This is a hypothetical cadence, not a measured speedup. If several such cycles run concurrently, the human cannot participate in every conversation. Multiple short feedback loops may therefore need to feed into a larger loop, moving the organization toward more continuous feedback than its familiar weekly rhythms allow.
Supervision could also become more fluid. A person who understands one domain might oversee its agents today and move to another project tomorrow. That flexibility requires training; it does not arrive automatically with the tools. As agents become more trustworthy and take on more tasks, Debois’s pyramid imagery moves the human portion upward into oversight and potentially makes it smaller.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Faster onboarding must still produce understanding
AI can help new developers join a team through tailored explanations and more personalized learning, reducing the interactions needed to get started. But ease of access to an answer raises a harder question: does the newcomer understand the system well enough to make a decision?
Supervision requires understanding. A person may no longer exercise detailed knowledge every day, but they still need it when deciding whether an output is acceptable or recovering from a failure. Access to a system with broad expertise does not eliminate that responsibility.
This complicates the claim that only seniors will have a place in AI-assisted teams. Debois suggests that personalized AI support and training could help juniors develop expertise faster. At the same time, the expectation of arriving as an instant specialist may discourage people from entering the field. Training programs would need to support that transition rather than assume the required judgment is already present.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From delegated services to autonomous peers
Companies already delegate work to SaaS providers. Delegating it instead to assistants could change company size and structure in the same way that shared assistants change teams. A further step comes when an agent performs a particular task as well as a human: it begins to occupy the position of a peer rather than a helper.
Debois connects this possibility to the working culture encoded in language models. Just as Wikipedia transmits knowledge, models can carry patterns of how people coordinate. In his agent-town example, simulated inhabitants interact with one another; he describes communication and collaboration as improving their results within that setting. The example illustrates social behavior among agents, not measured productivity in a business.
A useful peer would be able to go away and act on someone’s behalf. The limiting question is when that autonomy deserves trust. The same idea can be extended into a simulated software company, with agents playing investor, product-manager, and other roles. Debois finds the experiment interesting but explicitly says he would not trust the whole-company simulation at that point.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Agent teams inherit coordination problems
Smaller teams supported by agents might also need fewer leadership roles. But replacing people does not necessarily remove difficult team dynamics. An agent can exhibit unwanted behavior, and one harmful participant can influence others—the agent equivalent of a bad apple affecting a human team. The behavior of the group therefore matters alongside the capability of each member.
The consequences reach people outside the team. Debois describes a supermarket example in which AI controls entry and requires a person to smile. The inclusion question is concrete: what happens to someone who does not satisfy the behavior the system demands? It is an example of the authority delegated to a machine, not just the quality of its answers.
He compares OpenAI’s Model Spec to an organizational code of conduct: a statement of what a model should and should not do. The document describes intended behavior and guides training; it does not itself guarantee compliance. The analogy nevertheless helps separate having a capable agent from specifying acceptable conduct.
Debois does not expect these systems to acquire feelings. His concern about mistreating them is instead behavioral: humans can influence agents, just as agents can influence one another. Harmful conduct introduced through either route may spread through the team.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Specialists, employees, or managers?
An agent market could differentiate much like a human workforce:
- Generalists cover a broad range of tasks.
- Specialists provide stronger capability in a particular domain.
- Handcrafted agents address niche work for which no suitable agent is available.
Some might work continuously; others would be brought in for occasional specialist tasks. Their training and reward systems could differ accordingly.
Organizational records are a more contested extension. Debois describes a proposal to give agents employee records that was subsequently withdrawn, which he attributes to backlash. Representing an agent inside an HR system is distinct from treating it as human or granting it equal rights. The useful question is how an organization should account for these participants without confusing those categories.
Authority raises a separate question: could an agent manage people? A system deciding who enters a store already exercises a kind of managerial discretion. The possible positions now extend from assistant to peer to manager. Each expansion requires rules about decisions, codes of conduct, guardrails, and a human who understands and manages failures.
Once agents appear beside people in an org chart, staffing becomes a comparison of cost and output. Debois asks whether an agent might be equivalent to five humans; he offers no measured conversion ratio. The planning question is whether to hire someone or obtain the required output from an agent team. The slide makes this tangible by placing a legal-services agent alongside human roles in a Legal Affairs org chart.
The same shift could reach the labor market. Debois shows the idea of agents applying for LinkedIn jobs and imagines individuals managing their own agent task forces. A person would offer the work of that group, potentially changing how buyers find people, services, and capacity.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Evaluating agents and simulating the organization
If agents are managed as part of the workforce, they also need performance reviews. Are they producing value? Would different incentives improve their contribution? Individual, team, and business incentives operate at different levels, so steering an agent requires deciding which outcomes should shape its behavior.
Debois expects return-on-investment accounting to become more explicit for agents than it is for many human roles. Organizations may assess the return from a particular task or from an entire agent team, making performance evaluation part of deciding which work to delegate.
That leads to an organizational digital twin: a simulation in which a company could explore the effect of adding a person or an agent before changing the real team. An API might expose the semantics of the org chart and HR records to support such experiments. Debois treats that API as a thought experiment, with sensitive personnel information posing a substantial barrier.
The simulation and management machinery also creates another layer of authority. If people control agents, but agents increasingly allocate or supervise people’s work, who ultimately controls whom? Oversight itself becomes something that needs oversight.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Cheap agent labor could bring services back inside
The changes need not stop at the org chart. A conventional software business employs a team to build software and sells that software to customers. If agent work becomes sufficiently cheap, a customer that previously bought a third-party service might instead use internal agents to build and operate the systems it needs.
That changes the boundary between buying a service and doing the work internally. Debois describes the reversal as moving from software as a service toward “service as a software”: software increasingly performs the service itself. The economic condition matters—insourcing becomes attractive only if the agent-based alternative can provide the required service at an acceptable cost.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Build the system that does your work
The progression runs from individual assistance through team participation to agents performing work that changes the business itself. The closing strategy slide presents four levels: adding GPT, automating workflows, replacing workflows, and doing customers’ work. These are possibilities rather than a clear-cut path that every organization will follow.
Debois leaves the more futuristic imagery—hyperhumans and cyborg-like combinations—open. The outcome depends on how capable the technology becomes and how much trust it deserves. Organizational effects could arrive faster than expected, but their direction is not settled.
Even the familiar claim that a person using AI will replace someone who does not use it leaves larger questions unanswered. Removing toil could be beneficial, but how will people be paid, and how will the economy function as agents perform more work? Individual adoption alone does not settle those questions.
Debois’s practical advice is to move up a level: build the AI that builds the things you currently build. Knowing the work puts you in a position to specify it, recognize whether the result is good, and supervise the system performing it. Developing that capability creates a new source of value while preserving the understanding needed to take responsibility for the outcome.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
OpenAI’s original announcement explains the objectives, rules and defaults used to guide intended model behavior.
Outlier describes freelance expert work reviewing and improving AI outputs, including contributions from coding specialists.
Further reading
Conway’s April 1968 paper explains how communication structures constrain the systems organizations design.
A study of 25 agents in a simulated town, using memory, reflection and planning to produce believable individual and social behavior.
The July 2024 proposal to place digital workers in employee records and org charts, including the subsequent withdrawal notice.
Contemporary bookseller commentary on the 2011 multimillion-dollar Amazon listing and interacting automated pricing programs.
Full transcript
Loading transcript…