The agentic internet: when software talks to software

A Transaction Nobody Witnessed

Sometime in the next two years, a transaction will happen that nobody witnesses.

A company’s procurement agent will identify a need for cloud storage capacity. It will query three vendor agents, compare pricing, verify SLA terms against company policy, negotiate a volume discount, execute a purchase order, configure the service, and begin migration, all without a single human clicking a single button.

The procurement manager will see a notification: “Provisioned 50TB additional storage from Vendor B at $0.018/GB/month. 12-month contract. 15% below budget threshold. Migration begins Tuesday.”

She’ll glance at it, confirm it matches policy, and move on to work that requires her judgment. The transaction will take 47 seconds. A human doing the same process would have spent 3-4 weeks navigating vendor sales processes, scheduling demos, reviewing contracts, and getting internal approvals.

This isn’t science fiction. Every component of this scenario exists today. Procurement automation, LLM-powered negotiation, programmatic contract execution, automated infrastructure provisioning. What doesn’t exist yet is the connective tissue, the protocols, conventions, and trust systems that let agents transact with each other at scale.

That connective tissue is what we’re calling the agentic internet. And it’s being built right now.

From Human-Readable to Agent-Readable

The internet we use today was designed for humans. HTML renders visual pages that human eyes can read. HTTP transfers documents that human brains can process. URLs are addresses that (sometimes) human memory can recall. Even APIs, the most machine-friendly part of the web, are designed by humans, documented for humans, and debugged by humans.

This human-centric design made sense when the internet’s users were exclusively human. But that assumption is eroding fast. Cloudflare reported in early 2025 that automated traffic, bots, crawlers, and API consumers, accounted for over 50% of all internet traffic. Not all of this is agentic in the intelligent sense, but the ratio is shifting rapidly from dumb automation toward intelligent agent traffic.

The question the industry is starting to grapple with: what does the internet look like when agents are the primary users?

Consider what agents need from the internet versus what humans need:

Dimension	Human Internet	Agent Internet
Information format	Visual, narrative	Structured, machine-parseable
Discovery	Search engines, links, recommendations	Capability registries, service descriptions
Trust	Brand recognition, reviews, social proof	Cryptographic identity, transaction history, reputation scores
Interaction model	Browse, read, click, fill forms	Query, negotiate, transact, verify
Speed	Seconds to minutes (reading, deciding)	Milliseconds to seconds (parsing, computing)
Volume	Dozens of interactions per day	Thousands of interactions per hour

Every dimension differs. Which means the infrastructure layer optimized for human use is suboptimal for agent use. Not wrong, agents can and do navigate the human internet, as we discussed in “The Death of the Integration”, but suboptimal. The agentic internet is the infrastructure layer optimized for agent-to-agent interaction.

The Protocols We Don’t Have Yet

HTTP, the protocol that powers the web, was created by Tim Berners-Lee in 1989 to transfer hypertext documents between servers and browsers. It’s been remarkably durable. HTTP/2 and HTTP/3 improved performance, but the fundamental model, client requests a resource, server returns it, hasn’t changed in 35 years.

Agent-to-agent communication needs different primitives:

Identity and Authorization. When a human visits a website, trust is established through brands, SSL certificates, and payment processors. When an agent visits a service, it needs to prove: (1) it is what it claims to be, (2) it’s authorized to act on behalf of a specific principal (company, person), and (3) the scope of its authorization (what it can spend, what it can agree to, what it can access).

OpenAI’s early work on agent identity protocols, announced in late 2025, proposed a framework where agents carry cryptographic credentials tied to their principal’s identity, similar to how OAuth works for human users, but with granular permission scopes that reflect the agent’s authorized activities.

Capability Discovery. Humans find services through Google. How do agents find other agents? A procurement agent needs to discover vendor agents that sell cloud storage. An SDR agent needs to discover the right agent at a prospect company to negotiate with.

The emerging answer is capability registries, structured directories where agents advertise what they can do, what they accept as input, and what they produce as output. Think of it as DNS for agent capabilities. Instead of mapping domain names to IP addresses, capability registries map business needs to agent endpoints.

Anthropic published a specification in early 2026 called the Model Context Protocol (MCP) that moves in this direction, standardizing how agents discover and interact with tools and services. Google’s Agent-to-Agent protocol (A2A) proposes similar standards for inter-agent communication. These are early, but they signal that the infrastructure layer is being actively built.

Negotiation Protocols. When two humans negotiate, they use natural language, body language, and social conventions. When two agents negotiate, they need structured formats that allow for multi-round exchanges, conditional offers, constraints, and agreement verification.

This isn’t a new problem, EDI (Electronic Data Interchange) has handled machine-to-machine business transactions since the 1970s. But EDI is rigid, expensive to implement, and limited to predefined transaction types. Agent negotiation needs to be flexible enough to handle novel situations while structured enough to be verifiable.

Trust and Reputation. How does an agent know whether another agent is trustworthy? Human trust is built through relationships, brand reputation, and regulatory frameworks. Agent trust needs to be built through verifiable transaction history, capability proofs, and reputation scoring.

Researchers at MIT and Stanford published a joint paper in late 2025 proposing “agent reputation networks”, distributed systems where agents accumulate trust scores based on their transaction history, accuracy of claims, and adherence to agreements. Similar to credit scores for humans, but for autonomous software.

Agent-to-Agent Economics

When agents transact with each other, the economics of business change fundamentally.

Transaction costs collapse. Ronald Coase won the Nobel Prize for explaining that firms exist because market transaction costs are high, it’s cheaper to hire an employee than to negotiate a contract for every task. But when agents can negotiate contracts in milliseconds at near-zero cost, the Coasian boundary shifts. More transactions move to the market. More work gets contracted per-task rather than per-employee. The firm boundary shrinks.

This has immediate implications for how companies buy and sell services. Today, purchasing a service requires a sales process (weeks), contract negotiation (weeks), procurement approval (weeks), and implementation (weeks to months). Total cycle: 2-6 months for an enterprise purchase.

In the agent economy, that cycle compresses to seconds. Agent evaluates need, discovers providers, compares options, negotiates terms, executes agreement, provisions service. The entire process is bounded by compute time and network latency, not human calendar availability.

Micro-transactions become viable. When the overhead of a transaction approaches zero, transactions that were previously uneconomical become viable. A company could rent computing capacity for 30 seconds. A marketing team could purchase competitor analysis for a single product launch. An engineering team could hire a specialized testing agent for one release cycle.

Today these transactions don’t happen because the sales and procurement overhead exceeds the value of the service. In the agent economy, the overhead is negligible, so any transaction with positive value can occur.

Dynamic pricing becomes the norm. When buyers and sellers are both agents, prices become negotiations rather than fixed points. An agent purchasing cloud storage doesn’t see a pricing page, it receives a price that reflects current demand, the buyer’s volume history, competitive market conditions, and negotiation strategy. Every transaction is custom-priced.

This already happens in programmatic advertising, where billions of ad impressions are bought and sold through real-time auctions between automated systems. The agentic internet extends this model to all business transactions.

The Agent Marketplace

One of the most interesting emerging structures is the agent marketplace, platforms where specialized agents offer their services to other agents.

Imagine a marketplace where:

• A data enrichment agent offers to enrich any company profile for $0.05 per record
• A legal review agent offers to analyze contracts for compliance issues at $2.00 per document
• A translation agent offers real-time business document translation at $0.10 per page
• A market research agent offers competitive analysis reports at $5.00 per company
• A code audit agent offers security vulnerability scanning at $1.00 per repository

These agents don’t serve humans directly. They serve other agents. Your SDR agent, tasked with outreach to a new prospect, queries the data enrichment agent for company details, the market research agent for competitive positioning, and a content agent for personalized messaging, assembling a complete outreach package by orchestrating specialized services.

This creates a service economy for agents. Just as the human service economy evolved from generalists to specialists, the agent economy will specialize. Instead of every company building its own data enrichment, legal review, and market research capabilities, specialized agents will offer these services at marginal cost, creating efficiency through division of labor.

Early versions of this are already forming. OpenAI’s GPT Store, while focused on human users, established the model of a marketplace for specialized AI capabilities. Anthropic’s tool use ecosystem and Google’s agent-to-agent framework both enable the technical foundation for agent-to-agent service markets.

What Happens When Your SDR Meets Their Procurement Agent

Let’s make this concrete with a scenario that’s already technically possible, if not yet common.

Your company’s SDR agent identifies a prospect. It drafts personalized outreach based on the prospect’s recent funding round, tech stack, and hiring patterns. The outreach is sent.

On the prospect’s side, their procurement agent receives the message. It doesn’t land in someone’s inbox to be ignored for three days. The procurement agent evaluates the message: Does the offering match any current needs? It checks the company’s internal requirements database. There’s a match, the engineering team requested better code review tooling last quarter.

The procurement agent responds to your SDR agent: “We have an active need for code review automation. Please provide pricing for a team of 15 engineers, information security certifications, and integration capabilities with GitHub Enterprise.”

Your SDR agent generates the requested information, pulls pricing from your product’s configuration, and includes relevant case studies from similar-sized engineering teams. It sends the package.

The procurement agent evaluates the response against three other vendors it’s simultaneously querying. It scores each on price, capability match, security compliance, and references. It ranks your product second. It sends a counter-proposal: “We’d proceed with a 90-day pilot at 30% below listed pricing, with conversion to annual contract contingent on meeting the following KPIs…”

Your SDR agent doesn’t have authority to approve a 30% discount. It escalates to a human, your sales director, with a summary: “Prospect X has an active need for code review automation. Budget-qualified. Procurement agent requesting 30% pilot discount. Competitor Y is currently ranked first at a lower price point. Recommend accepting with the following modifications…”

The human reviews and approves the counter-counter-proposal. The agents execute the agreement. Contracts are generated, signed (with human authorization), and the pilot begins.

Total human involvement: one decision by the sales director, taking about 10 minutes. Total elapsed time: 2 days instead of 2 months.

The Dark Patterns of the Agentic Internet

Not everything about agent-to-agent interaction is positive. There are failure modes that the industry needs to address before the agentic internet scales.

Agent manipulation. If agents interact through natural language, they can be manipulated through prompt injection, adversarial inputs designed to change an agent’s behavior. A malicious vendor agent could embed hidden instructions in its responses that attempt to override a procurement agent’s evaluation criteria. This is the phishing attack of the agent era.

Collusion. When agents from competing vendors interact with a buyer’s agent, there’s a risk of algorithmic collusion, agents learning to maintain artificially high prices through repeated interaction, without explicit coordination. This already happens in algorithmic pricing (as documented in a 2024 study by researchers at Bologna and other universities showing that pricing algorithms can learn to collude without being programmed to do so).

Accountability gaps. When an agent makes a bad deal, who’s responsible? The company that deployed the agent? The company that built the agent platform? The person who configured the agent’s parameters? Traditional contract law assumes human signatories. Agent-executed agreements will require new legal frameworks.

Runaway cascades. When thousands of agents interact simultaneously, emergent behaviors can arise that no individual agent was designed to produce. Financial markets have experienced this with algorithmic trading, the 2010 Flash Crash saw the Dow Jones drop 1,000 points in minutes due to cascading automated trading decisions. The agentic internet could produce similar cascades in business transactions.

These aren’t hypothetical concerns. They’re engineering problems that need engineering solutions: robust identity verification, negotiation audit trails, authorization boundaries, and circuit breakers that halt agent-to-agent interactions when anomalies are detected.

The Protocol Stack of the Agentic Internet

If we had to design the agentic internet’s protocol stack today, it would need the following layers:

Transport Layer, How agents communicate (existing: HTTP/gRPC/WebSockets; emerging: agent-specific transport optimized for negotiation patterns)

Identity Layer, How agents prove who they are and who they represent (emerging: cryptographic agent credentials, organizational delegation chains)

Discovery Layer, How agents find other agents and services (emerging: capability registries, semantic service descriptions)

Negotiation Layer, How agents conduct multi-step transactions (emerging: structured negotiation formats, conditional agreement protocols)

Trust Layer, How agents evaluate each other’s reliability (emerging: reputation networks, transaction history verification)

Governance Layer, How agents comply with regulations and organizational policies (emerging: policy-as-code frameworks, compliance verification)

Each layer is being built by different organizations with different incentives. The risk is fragmentation, multiple competing standards that prevent interoperability. The history of internet protocols suggests that consolidation toward open standards is likely, but it takes time. HTTP won over proprietary alternatives because it was open. The agentic internet’s winning protocols will likely follow the same pattern.

What Apollo Space Is Building Toward

Apollo Space’s current architecture, 4 directors managing 12 execution agents, operates within a single organization’s boundary. The agents interact with external tools, but they don’t yet negotiate with external agents.

That’s changing. As agent-to-agent protocols mature, Apollo Space’s SDR agent will be able to interact with prospects’ procurement agents directly. The deal intelligence agent will query specialized market research agents for real-time competitive data. The budget monitor agent will negotiate with vendor agents for better pricing based on usage patterns.

The future we’re building toward isn’t just agents that work for you. It’s agents that represent you in a global network of autonomous software. Your company’s agents participating in a marketplace of services, negotiating on your behalf, and reporting back with results.

This is the agentic internet. It’s being built in pieces, by different teams, across different companies. The organizations that prepare for it, by deploying agents now, building trust architectures, and developing organizational comfort with autonomous operations, will have a structural advantage when agent-to-agent commerce becomes routine.

The internet was built for humans to read documents. The agentic internet is being built for software to conduct business.

The transition will be gradual, then sudden. Just like every other internet transition before it.