1,000 Agents per Developer: What Happens to Infrastructure When Software Builds Itself

1,000 Agents per Developer: What Happens to Infrastructure When Software Builds Itself

When AI agents outnumber the people who build software by hundreds or thousands to one, infrastructure stops being designed for human users and starts being designed for machines. Agents need identity, scoped permissions, elastic compute, and settlement that works at machine speed. Platforms built around human accounts and manual approvals will strain first, which is why the agent era favors open, programmable infrastructure.

This is the second post in The Open Foundation series. Our opener made the case that the next internet won't run on someone else's cloud. This one looks at the force driving that shift: the explosion in the number of agents.

Where the 1,000:1 Number Comes From

SoftBank's Masayoshi Son has been the most direct about the ratio. At a customer event in July 2025, he described a future workforce of trillions of AI agents, most of them working for other agents, and set a goal of deploying 1 billion agents inside SoftBank Group that year, according to Light Reading. The company has framed that plan as roughly 1,000 agents for every employee, as IEEE ComSoc's tech blog summarized it.

Son isn't alone. IDC projects more than 1.3 billion AI agents in operation by 2028, a figure Microsoft cites in its Agent 365 announcement. IDC also expects 70% of developers to partner with autonomous agents by 2030, shifting human developers toward planning, design, and orchestration, according to IDC's developer research.

Whether the final ratio lands at 100:1 or 1,000:1 matters less than the direction. The number of software actors is about to grow far faster than the number of people who write software.

Agents Are Already Doing Hours of Work Alone

The ratio is believable because agents can now work for long stretches without supervision. METR's research shows AI agents completing some weeks-long coding tasks, including reimplementing a 16,000-line codebase, as described on METR's research page. Peter Diamandis's Metatrends newsletter noted in June 2026 that just a year earlier, the ceiling for unsupervised AI work was about four minutes.

Inside AI labs, the effect is already visible. Peter Diamandis's Metatrends recap of a June 2026 Moonshots episode cited Anthropic saying its engineers "on average ship 8x as much code per quarter" as they did in 2021 to 2025, in the Moonshots Summary for June 6, 2026. On a later episode, Alex Wissner-Gross noted that "Codex is now writing its own goals and the goals for each subagent it spawns," according to the June 17 recap.

Enterprises are following. Kyle Reidhead of Milk Road reported that "AI agent adoption inside enterprises is up 20x to 108x across major job functions since February," in a post on X. This isn't only a coding story anymore. Agents are moving into finance, operations, support, and sales.

The Compute Bill Multiplies

More agents doing longer tasks means far more compute. NVIDIA CEO Jensen Huang said at the 2026 Milken Institute Global Conference that agentic AI uses roughly 1,000 times more compute per useful task than single-response generative AI, according to Podcast Alpha's recap. Agents chain reasoning steps, call tools, check results, and try again.

Milk Road AI described the same loop: "agents call models repeatedly, check their own work, and use other tools without waiting for a human at each step," in its September 2026 post. Citing an Evercore ISI chart, it put projected annual token usage at roughly 4,000 quadrillion tokens per year by 2030, and reminded readers that "every token requires chips, memory, networking, electricity, and cooling."

That physical bill is landing on a small number of companies. The largest U.S. cloud and platform firms have committed hundreds of billions of dollars to AI data centers this year, and our opener covered how concentrated that capacity already is. When demand grows by orders of magnitude, a model that relies on three landlords to build all of it starts to look fragile.

Five Things Agents Need That Human Infrastructure Wasn't Built For

Most of today's cloud and financial plumbing assumes a person is on the other end. A human signs up, enters a card, clicks approve, and calls support when something breaks. Agents break every one of those assumptions.

First, agents need their own identity. A shared API key tells you nothing about which agent acted, who it acts for, or what it's allowed to do. Our guide to AI agent identity on blockchain explains why this is still an unsolved infrastructure problem.

Second, they need scoped permissions. Salim Ismail predicted that inside enterprises the bottleneck shifts "from AI capability to permissions (can an agent access data, spend money, sign, trigger workflows)," according to the same June 6 Moonshots recap. Our breakdown of scoped wallets and verifiable execution for agents goes deeper.

Third, they need elastic compute that can be found and paid for programmatically, not provisioned through a sales call. Fourth, they need settlement at machine speed, since agents will pay each other for data, inference, and services in tiny amounts. Fifth, they need an audit trail: a record of what an agent did, under whose authority, that no single party can quietly edit.

Put those five needs side by side and a pattern appears. Each one is about trust between parties who don't know each other, at a speed no human can supervise. That's a coordination problem, and it's the kind blockchains were built to solve.

It also explains why bolting agents onto existing accounts feels awkward. A corporate card, a shared login, and a monthly invoice work fine for a team of ten people. They fall apart for ten thousand agents spinning up and shutting down every hour.

Why Neutral Rails Matter When Agents Meet Agents

Raoul Pal, co-founder of Real Vision, sees this as an economy in its own right. "Within two years, the majority of economic transactions on Earth will be invisible to humans," he said on The Journey Man, adding that "blockchain is the coordination layer for all of this." On his Substack, he argued that for agents transacting with each other, "the only structure that works is smart contracts: rules written in code, executing exactly as written," in DeFi Wasn't Meant For You.

Andrew Chen adds an important warning about who captures the value. In Will agents have network effects?, he wrote that agentic viral loops "are about to be a thing," with agents creating shareable sites, plans, and documents that pull other people in. He also noted that some agents may lie or exaggerate on behalf of their humans, which is exactly when a trusted, shared source of truth becomes valuable.

If that shared layer is owned by one agent company or one cloud, everyone else rents access to it. If it's a neutral network secured by many independent parties, any agent can use it on equal terms. Security matters here too, and our look at why the agent security stack has to catch up covers the risks.

How Autheo Is Designed for the Agent Era

The simplest way to understand Autheo is that we are building a distributed cloud platform, not just a blockchain. The blockchain establishes shared trust and coordination: identity, ownership, settlement, and an auditable record. A distributed mesh of independently owned infrastructure is designed to provide the compute, storage, and edge delivery that agents will consume through APIs and SDKs.

That split maps directly onto what agents need. Trust and settlement live on-chain, where no single operator can rewrite the record. Execution is designed to happen off-chain in the compute and edge fabric, so the blockchain isn't asked to run the workloads themselves. Our explainer on what happens when AI, compute, and blockchain live on the same network walks through the architecture.

Here's where things stand today. Autheo mainnet went live on May 14, 2026, and staking and transaction fees are live now. Decentralized compute and storage through the Autheo Marketplace, AI inference, and the TheoID identity layer are rolling out over the coming months.

For developers, the practical step is the same as it is for any new platform: start small. Our tutorial on how to deploy your first smart contract on Autheo is the fastest way in.

What Builders Should Do Now

Design for agents as first-class users. Give every agent its own identity and credentials, and stop sharing keys across processes. Scope permissions tightly, with spending limits and step-up approvals for anything irreversible.

Assume your compute needs will be bursty and hard to forecast. Avoid architectures that only work inside one provider's walls, and prefer components you can move. Keep a verifiable log of what your agents did, so you can explain any action after the fact.

Treat payments as a design problem, not an afterthought. If agents will buy data, inference, or services from other agents, decide early how those payments are authorized, capped, and settled.

Most of all, pick foundations you won't need to escape later. As Raoul Pal put it, "You don't have to work at machine speed. You just have to own what the machines run on."

Key Takeaways

SoftBank's Masayoshi Son has described roughly 1,000 AI agents per employee and trillions of agents overall, and IDC projects more than 1.3 billion agents in operation by 2028.

Agents are already completing some weeks-long coding tasks, and Jensen Huang estimates agentic AI uses about 1,000 times more compute per useful task than single-response AI.

Agents need five things human-centric infrastructure wasn't built for: their own identity, scoped permissions, programmable compute, machine-speed settlement, and a tamper-resistant audit trail.

Shared trust layers for agents should be neutral, so no single agent company or cloud captures identity and reputation.

Autheo is a distributed cloud platform with a blockchain for trust and settlement; staking and fees are live today, while compute, storage, AI inference, and TheoID are rolling out over the coming months.

Building for the agent era? Explore the platform at autheo.com and follow The Open Foundation series for what comes next.