Agent Economics: Who keeps the ledger of AI Labor?

Brayden Yancy — Tue, 20 Oct 2026 22:12:52 +0000

Prologue:

It’s no secret that software today is more capable than it has ever been. With the emergence of increasingly powerful artificial intelligence and the multitude of designed input options exposed to individuals, humans can prompt software into existence in the blink of an eye.

This shift is forcing a change in our perspective on how to handle the widely adopted and powerful modern AI. Currently sitting at the top of the mountain are the monolithic enterprises OpenAI, Anthropic, Google, xAI, and Meta. As capitalism strives to win in this fast-forwarded geo-political arms race, there are too many ways things can turn out, not all of which are accounted for.

This emerging system is increasingly described as an ‘agent economy’ but the economics underneath it, like what the agents cost, what they produce, and what they exchange is what is referred to as “Agent Economics”.

Agent Capabilities

“Agent Economy” shows up in research papers, keynotes, and company blogs across the internet, but is rarely defined. So, before getting into who is doing what, it’s important that we stop and try to understand the entire picture. To do this, you must first understand from the ground up, starting with: What is an AI agent?

With the emergence of massive LLMs, an old problem became newly important: chaining steps together. A chatbot could return information en masse but lacked the ability to complete actual tasks that required more than a direct response in its own environment. This mimicry of human communication was impressive, but for the front runners it was only the beginning. How can we create the necessary infrastructure to allow this single environment to expand into others and complete tasks beyond the current box? Modern AI agents solve this issue by giving models the ability to act beyond a single response, and work outside of its box.

Now capable of planning, calling tools, reading results and making decisions until the desired goal is reached, we have created the necessary extensions for an AI ‘brain’ to interact with things beyond its current environment.

This idea is older than today’s models. Researchers such as Pattie Maes at the MIT Media Lab were studying "software agents" in the 1990s: programs that act on a person's behalf. What changed is capability. Modern models can carry out long, multi-step work well enough that companies can now hand them real tasks. So, where do we go from here?

Economic Questions

Now that software is capable of acting on someone’s or something's behalf, many large questions quickly follow.

Who is responsible for an agent's mishaps?
What is the cost to complete particular tasks? What tasks specifically?
How do we monitor actual work and actions done by the agents?
Who leads this initiative?

We can speculate on who is responsible for an agent’s actions but, in my opinion, that comes down to where we end up in our new “Agent forward economy”. So to answer that question I raise another - What should the goal be?

More powerful and capable agents are created daily by large corporations, but is this the path we expect to funnel our resources into until we hit a wall? Well, only if we continue to build “mothership” AI.

Project NANDA has its own initiative aimed at creating an open infrastructure layer where agents can find, verify and coordinate with one another. Agents are already specialized into individual roles and work together as a group to complete larger tasks, but where do they live? OpenAI’s servers? Shouldn’t we treat AI like DNS, or the Internet? Ramesh Raskar, MIT Media Lab leads Project NANDA and pictures billions or even trillions of agents, one or more for every person. In a July 2026 MIT Sloan piece, he argues the big opportunity is not building agents for specific jobs but building what agents need.

“The economy of the future is not an economy where we say, ‘Let’s create agents for X,’” he said. “It’s an economy where we instead say, ‘Let’s create X for agents.’”

So, now we have two possible paths: large mothership AI with fleets of smaller agents, or Ramesh’s Oprah approach where, “You get an agent! AND you get an agent!”.

However, this is the real world, so we cannot simply expect one of these ideas (not to mention the many others actively being pursued) to be the only potential outcomes.

Google DeepMind has another outlook. In Virtual Agent Economies (a 2025 preprint), Nenad Tomašev and colleagues propose the "sandbox economy." They map agent economies along two lines: did the economy emerge on its own or was it designed, and how freely does value flow between it and the human economy? Their question is what should the market's rules be? They suggest tools such as auctions for fair allocation and "mission economies" organized around shared goals.

This starts to get hectic pretty fast though, so let’s ground ourselves for a moment. Across these different visions, the same infrastructure problems keep appearing: how do we codify things like identity, trust, coordination, accountability and eventually payments.

What we don't know

It is worth saying plainly, none of this is settled. The agent economy described above is a set of competing visions, not a description of today.

Nobody really knows the timeline. I mean Raskar himself says it could take two years, five or even ten. The “Agent Economy” could stay open, like the early web, or consolidate under a few large companies. Raskar thinks this is far more likely (“nine out of 10 paths,” he says). It might be carefully designed, as the DeepMind authors hope, or simply pile up as adoption outruns planning.

Today, organizations are paying for agent work and many cannot yet say what they get for it. In my mind, I think we should be treating AI Agents like the workers they are and actually keep a ledger and track record of their work. Of all these questions, one is already becoming practical. How do we keep a record of the work agents are doing today?

From AI consumption to AI labor

In my opinion, the missing layer is attribution: connecting the resources an agent consumes to the work it actually produces.

Software gives us a great place to start. Coding agents leave records of their usage, while the work itself leaves another record in Git. The Git history contains commits, branches, pull requests and releases. By connecting those records together, we make it possible to track beyond ‘how much AI did we use?’ toward a new grounded view of, ‘what did the AI help us accomplish, and what did that work cost us?’

The meaningful unit is the completed work those commits belong to: the feature, fix, release or other outcome.

That is what measuring AI labor should move toward:

AI usage → attributable work → completed outcome → economic cost

Over time, the same idea could extend wherever agents work and leave an observable result: a support case resolved, a research report delivered, a campaign launched or a financial analysis completed.

The exact unit will differ by industry, the principle will not.

If agents are going to participate in an economy, somebody has to keep the ledger; not only of what they consumed, but of what they did.

Where Centrail fits

The ledger specifically is the problem Centrail is starting with. Today, Centrail connects coding-agent usage to the Git history where its work appears, creating a ground level record of AI usage and the software being produced alongside it.

The larger goal is a system of record for AI labor. A way to understand what agents consumed, the work they performed and, ultimately, the outcomes they produced.

Whether the agent economy becomes open or consolidated, designed or emergent, those questions remain the same. If software is going to do more of the work, we are going to need a better way to account for the work it does.

This is the first piece in an ongoing series of agent economics.

Sources & Further Reading

I tracked 2.86 billion AI coding tokens. Here's what a commit actually costs.

Brayden Yancy — Thu, 24 Sep 2026 17:46:04 +0000

Every team I talk to is paying for Claude Code, Copilot, Codex or all three. Almost none of them can answer a simple question: what did that AI spend actually ship?

The provider invoice tells you how many tokens you bought. It doesn't tell you which commits those tokens turned into. So I built a tool that does, and pointed it at my own work first.

The numbers

Here's everything Centrail has tracked on my own AI-assisted development so far:

Metric	Total
Tokens	2.86 billion
AI spend	$4,099
Commits	944
Cost per commit	$4.34

Why cost per commit matters

"We spend $X a month on AI" is a budget line. "$4.34 per shipped commit" is a unit cost you can actually manage. Once you have it, you can ask better questions:

Which repos burn the most tokens per commit?
Which models cost more without shipping more?
When did spend spike, and did output spike with it?
Are there cheaper models that would do the same job?

How it works

Centrail runs where the work happens, on your machine, not at the invoice level. The CLI reads local usage from your AI coding agents and ties each token and dollar to the Git commit it produced. That syncs to a dashboard, so the whole team or organization stays up to date on real AI metrics automatically. Nobody has to update a spreadsheet.

It also estimates the energy, water and CO2 behind every token, with a public methodology, because AI's footprint is a cost too.

Try it

The CLI is open source: github.com/braydenyancy/centrail-cli
Dashboard and free tier: centrail.org

What does a commit cost on your team? If you don't know, I'd love to hear how you're thinking about AI spend in the comments.

DEV Community: Brayden Yancy