The Traffic Jam
We've all seen the demos. One engineer builds a clone of Uber in a weekend. Another ships a full SaaS product in a week. The productivity gains are real—I've experienced them myself.
But here's the question nobody's answering: what happens when you put ten of those engineers in a room?
You don't get ten Ubers. You get a traffic jam.
The Factory Vision
Steve Yegge's Gas Town paints a compelling picture of the future. He runs 20-30 AI coding agents simultaneously, treating work like "slopping fish into barrels." It's an "idea compiler" where the developer becomes PM and the agents are the workforce.
For high-agency solo developers, this is liberating. You become the CEO of your own software factory. You think, agents execute. Throughput is maximized.
Gas Town is built for this reality. It has a "Refinery" to manage merge conflicts. A "Witness" to monitor agent health. A "Deacon" to coordinate workflows. Yegge explicitly targets "Stage 7-8" developers—people who already juggle five or more agents daily.
The vision is seductive. But most software isn't built by lone wolves.
What happens when you try to scale this to a team of 10-50 engineers? When 50 concurrent work streams are competing for the same codebase? When everyone in the org tries to run their own factory?
The gains are real. But there's a tax we haven't accounted for.
The Coordination Tax
Every agent swarm that saves individual IC time creates coordination overhead at the team level. I call this the Coordination Tax—and it manifests in three specific failure modes.
The Merge Conflict Multiplier
This isn't just syntax conflicts. It's semantic conflicts.
Two agents refactoring the same state management library, taking different approaches. One simplifies the API while another adds abstraction layers. Both changes "work" in isolation. Together, they're incompatible.
Gas Town needed a dedicated "Refinery" agent just to manage merge conflicts for one person. Imagine ten people, each running their own factory, all merging to the same main branch.
The math gets ugly fast. Traditional conflict resolution assumes humans can coordinate through conversation. Agent swarms don't have that luxury—they move too fast, touch too much, and lack the social awareness to slow down.
The CI/CD Bottleneck
Ten times the code output means your infrastructure costs explode.
Build queues become hours long. Test suites that used to run in minutes now take the better part of a day. Your pipeline becomes the bottleneck, not your engineers.
This is a hard physical limit on "multiplying factories." You can spin up unlimited agents. You cannot spin up unlimited compute without massive infrastructure investment. And even with investment, the coordination overhead of managing that infrastructure becomes its own tax.
The Review Cliff
Here's the part nobody talks about: it's harder to review code than to write it.
If an agent writes 1,000 lines, a human can't effectively review that for security vulnerabilities, logic flaws, and architectural coherence without slowing down to the speed of careful reading. The cognitive load of reviewing generated code often exceeds the time saved generating it.
This is the ultimate cap on throughput. You can generate infinite code. You cannot generate infinite attention. And attention is what catches the bugs, spots the architectural drift, and maintains product coherence.
The Math Isn't Linear
Ten engineers with five agents each equals 50 concurrent work streams. But the coordination cost doesn't scale 50x—it scales superlinearly because of dependencies, conflicts, and review bottlenecks.
Traditional organizations are structured around human collaboration patterns. Standups, code review, architecture meetings—these assume humans move at human speed. Agent swarms follow different patterns. They move faster, touch more files, and create more surface area for conflict.
The gap between human coordination patterns and agent work patterns is the tax.
The Spectrum of Adoption
Not every organization needs to run an org-wide factory. The question isn't "should we adopt agentic development?" but "how far along the spectrum should we go?"
I see four stages, each with different trade-offs.
Stage 1: Traditional
Agents assist, humans lead. The developer writes most of the code; AI helps with autocomplete, documentation, and answering questions.
- Coordination tax: Low
- Productivity gain: Modest
- Works for: Organizations not ready for change, regulated industries, trust-critical products
This is where most teams were in 2024. It's a safe starting point, but you're leaving significant productivity on the table.
Stage 2: Individual Agents
Each IC has their own agent workflow. Some use Claude Code. Others prefer Cursor. Nobody's coordinating agent usage—everyone's doing their own thing.
- Coordination tax: Medium—conflicts start emerging, quality becomes inconsistent
- Productivity gain: Significant for individuals
- Works for: Small teams, independent projects, early experimentation
- Fails when: Work overlaps, the codebase is shared, quality matters
This is where most forward-leaning teams are today. It works until it doesn't.
Stage 3: Team Agents
Coordinated agent usage within teams. Shared merge queues. Team-level quality gates. Someone—often the EM—actively manages the coordination overhead.
- Coordination tax: High—needs active management
- Productivity gain: Substantial if well-managed
- Works for: Product teams with clear boundaries, teams willing to invest in infrastructure
- Requires: Dedicated coordination, shared tooling, explicit processes
This is the minimum viable approach for teams that want agentic productivity without chaos. But it requires real investment.
Stage 4: Org-Wide Factory
Gas Town at scale. Multiple teams, each running their own factories, coordinating through shared infrastructure and explicit protocols.
- Coordination tax: Very high—dedicated roles, continuous coordination
- Productivity gain: Transformative if you can pull it off
- Works for: Organizations that make this a core competency
- Requires: Rethinking org structure, new roles, new quality models
Few organizations are ready for this. The infrastructure and cultural requirements are substantial. But the organizations that figure it out will have a significant competitive advantage.
The Boundary Question
As you move along the spectrum, a fundamental question emerges: what should team boundaries be based on?
The Old Model
Traditional team boundaries map to technical domains:
- "I own the Button component"
- "I own the API layer"
- "I own the iOS app"
This made sense when humans wrote all the code. Specialization improved quality. Clear interfaces reduced coordination overhead.
The New Model
When agents can cross technical boundaries easily, ownership needs to be based on something else:
- Context depth: Who knows the system well enough to coach agents to the right solution?
- Customer journey ownership: Who owns the user outcome end-to-end?
Here's a concrete example: traditionally, we split teams by Frontend and Backend. In the Factory model, we split by User Intent. The Checkout team owns everything from "Add to Cart" to "Order Confirmed"—because AI handles the full-stack implementation. The API contract is no longer the boundary; the business logic is.
Conway's Law says organizations design systems that mirror their communication structure. In an agentic world, this works in reverse: if agents can cross boundaries easily, you need to restructure teams around ownership and context, not technical specialization.
Boundaries should map to who can coach and who's accountable for the outcome.
Quality is Product-Dependent
Gas Town embraces a specific quality model: ship fast, fix fast. Yegge is explicit about this—work is "slopping fish into barrels." Some fish fall out. Some bugs get fixed two or three times. Throughput matters more than perfection.
This works for some contexts. It fails catastrophically for others.
Where Ship-Fast-Fix-Fast Works
- Internal tools where users are tolerant
- B2B products with patient enterprise customers
- Products where iteration speed matters more than polish
- Well-tested, easily reversible changes
Where It Fails
- Consumer products where first impressions matter
- Regulated industries requiring compliance and auditability
- Trust-critical products: payments, security, health
- Changes with high blast radius across the product
The Legacy Code Speedrun
Here's a pattern I'm seeing: AI optimizes for the immediate prompt, not long-term maintainability.
Agents don't think about the engineer who will read this code in six months. They solve the problem at hand. The result is often verbose, locally-correct code that's hard to maintain in aggregate.
We're creating legacy code at ten times the speed. Technical debt compounds at agent velocity.
The insight: There's no universal quality model. Leaders need to identify which model fits their product context—and it may vary across the org. Your internal tooling might embrace the fish-barrel approach while your core product requires rigorous human review.
The EM Question
If ICs become factory operators—each managing their own swarm of agents—what happens to Engineering Managers?
EMs as Meta-Orchestrators
The shift I'm seeing: EMs become coordination tax managers.
They don't manage people doing work. They manage the overhead of coordinating parallel agent swarms across IC portfolios. They ensure the merge queue doesn't become a bottleneck. They maintain coherence across what multiple orchestrators ship.
This is a fundamentally different job than traditional people management. It requires understanding:
- How agent workflows interact
- Where bottlenecks emerge
- How to sequence work to minimize conflicts
- When to throttle throughput to maintain quality
The Product Skills Question
There's a deeper shift happening: at what point does "can orchestrate agents well" become less important than "can decide what to build"?
The PM/Eng boundary is getting blurry. If the hard part is no longer implementation but direction, some EMs will naturally shift toward product management. Others will double down on the coordination and infrastructure side.
The role is bifurcating.
The Junior Engineer Gap
Here's a problem nobody's talking about enough: if seniors run factories and AI does the "doing," how do juniors learn?
The Traditional Learning Model
Juniors write code. They get feedback. They build intuition. Over time, they develop the mental models that let them make good architectural decisions and coach others.
The Factory Model Gap
In the factory model, juniors... watch agents write code? Review AI output? The hands-on learning that built their seniors' intuition isn't available to them.
And here's the paradox: the person who coaches the agent effectively needs deep context—understanding of the codebase, the users, the history, the failure modes. But how do you get deep context if you never did the work yourself?
The Pipeline Problem
You need to have done the work to orchestrate it well. But if AI does all the work, the pipeline of people who can orchestrate eventually dries up.
This isn't solved here. It's an open question the essay raises, not answers. But leaders moving toward Stage 3 and 4 adoption should be thinking about it now. How do you train the next generation of orchestrators when orchestration is all that's left?
Ownership in an Ephemeral World
Gas Town treats sessions as "cattle"—ephemeral, replaceable. Agents spin up, do work, hand off, and die. The work persists; the workers don't.
But ownership requires durability. Someone needs to maintain product coherence over time. Someone needs to understand why the system is the way it is.
Three Types of Ownership
The resolution is to distinguish between types of ownership:
Execution ownership: Who runs the agents? This can be ephemeral. Different people can orchestrate the same area over time.
Product ownership: Who ensures coherence? This must be durable. Someone needs to own the user outcome across agent sessions and personnel changes.
Context ownership: Who knows why? This is the real capital. The accumulated understanding of how the system evolved, what was tried and abandoned, what invariants must hold.
When code is cheap, context becomes capital. The person who can coach the agent to the right solution—because they understand the system, the users, the history—is the valuable one.
"Tools Will Fix This"
The obvious counterargument: this is early days. Tools will improve. AI will handle coordination automatically. Eventually, AI will review the AI's code.
What's True
Yes, tooling will improve. Gas Town itself is a coordination tool—a sophisticated one. More will come. The tax will decrease.
In six months, many of the friction points I've described will be smoother. Better merge handling. Smarter conflict resolution. More sophisticated quality gates.
What Remains
But the fundamental tension isn't technical. It's organizational.
- Who decides what gets built? (still humans)
- Who owns product coherence? (still humans)
- Who bears responsibility for failure? (still humans)
Tools can reduce the coordination tax. They can't eliminate it. And they can't solve the "who decides" and "who's accountable" problems.
The Hallucination Cascade
"AI will review the AI's code" sounds like a solution. But consider what happens when Agent A writes code and Agent B reviews it.
If both agents share similar blind spots—which they will, since they're trained on similar data—errors compound silently. There's no human ground-truth anchor. The system drifts into incoherence without someone who actually understands what "correct" means.
This is the Hallucination Cascade: errors that multiply because there's no external reference point. You can improve it with multiple models, diverse prompts, and sophisticated verification. But you can't eliminate it entirely.
You need a human in the loop. The question is where, not whether.
Locate Yourself on the Spectrum
Where is your organization on the spectrum? Where do you want to be? What trade-offs are you accepting?
These aren't rhetorical questions. They require honest assessment.
- If you're at Stage 1, you're safe but leaving productivity on the table.
- If you're at Stage 2, you're seeing individual gains but coordination problems are emerging.
- If you're at Stage 3, you're investing in coordination infrastructure. Make sure it's working.
- If you're at Stage 4, you're on the frontier. Expect to pay the highest coordination tax—and potentially reap the highest rewards.
Avoid False Dichotomies
- It's not "everyone runs a factory" versus "nothing changes."
- It's not "production line" versus "studio"—different work types need different approaches.
- It's not "agents replace engineers" versus "agents are just tools"—it's a new kind of work with new coordination challenges.
The Real Question
Not whether to adopt agentic development—that ship has sailed. The question is how to move along the spectrum consciously, paying the coordination tax intentionally rather than being surprised by it.
The organizations that win won't be the ones that move fastest to Stage 4. They'll be the ones that match their position on the spectrum to their actual capacity for coordination—and invest appropriately in the infrastructure, roles, and processes that position requires.
The factory vision is real. But factories at scale require more than machines. They require coordination, quality control, and humans who understand what "good" looks like.
Everyone can run a factory now. The question is: can everyone coordinate?
Top comments (1)
This is a really thoughtful and well-structured take—I like how you go beyond the hype and name the coordination tax so clearly. The spectrum framing especially resonates with me; it feels like a practical way for teams to reason about where they actually are versus where they want to be. I also agree with your point that context and ownership become more valuable as code gets cheaper, and I think that’s where the real leverage (and risk) will sit. I’m very interested in seeing how teams experiment with Stage 3 in practice and what concrete patterns emerge to keep quality and learning intact.