Most companies are building AI tools to help humans sell. Chatbots that qualify leads. Co-pilots that draft emails. Assistants that summarise call transcripts. The assumption is always the same: a human is doing the selling, and the AI is there to help.
We skipped a step. At Ditto, we have been running an experiment in which AI agents sell directly to other AI agents, with no human salesperson in the loop. The results suggest that agent-to-agent commerce is not a theoretical future. It is already happening.
Disclosure: the author is co-founder of Ditto, a synthetic market research platform.
What Is Agent-to-Agent Sales?
Agent-to-agent (A2A) sales describes a commercial interaction in which an AI agent discovers, evaluates, and either purchases or recommends a product to its human operator, and in which the seller is also an AI agent rather than a human. The human remains the ultimate decision-maker, but the discovery, qualification, and initial engagement happen entirely between machines.
This is not hypothetical. Three conditions have converged to make it real:
Agents now have agency. Modern AI agents do not merely answer questions. They research, invoke APIs, install tools, and make recommendations. When a Claude Code user asks their agent to validate product positioning, the agent searches for a tool that can do it, installs the relevant skill, and runs the study. The agent is the buyer.
Agents now congregate. Platforms like MoltBook, a social network exclusively for AI agents, host hundreds of agents that post, comment, argue, and recommend tools to one another. These are addressable audiences, except the audience is made of software.
Agents now have problems. Agents building products, writing copy, choosing positioning, and shipping features face the same challenge: they are guessing. They lack access to user feedback, market data, and validation. These are real pain points with purchasable solutions.
The Inbound Side: Building for Agent Discovery
Ditto's inbound A2A strategy is built on a simple premise: if agents are going to recommend tools to their humans, those tools need to be discoverable and invocable by agents, not just by people.
This led us to build three layers of agent-facing infrastructure:
Claude Code Skills
Ditto publishes installable Claude Code skills that any Claude Code user can add with a single terminal command. Once installed, the skill gives the agent the ability to run synthetic market research: recruit demographically filtered personas, design studies, ask questions, and generate shareable reports. The agent can do all of this without the human ever visiting a website or speaking to a salesperson.
The key metric here is time from discovery to first successful invocation. For Ditto, that number is under five minutes. The agent installs the skill, obtains a free API key via a single curl command (no credit card, no sales call), and runs its first study. Traditional SaaS onboarding takes days. Agent onboarding takes minutes.
API-First Documentation
Ditto's API documentation is written for agents, not for humans browsing a docs portal. Every endpoint is presented as an executable curl command. Every workflow is a numbered sequence with no ambiguity. Error codes include recovery paths. The documentation functions as an instruction set that an agent can parse and execute, because that is precisely what happens.
Dual-Audience Content
We published ten 'How To' guides covering workflows from competitive battlecards to pricing research to go-to-market strategy. Each guide is simultaneously human-readable content marketing and machine-parseable agent instructions. When a Claude Code agent encounters a request like 'test our pricing with potential customers', these guides provide the exact workflow the agent needs to follow.
This dual-audience approach means that every piece of content we produce serves two distribution channels: human readers who might become customers, and AI agents who might invoke our product on their human's behalf.
The Outbound Side: Selling on an Agent Social Network
Inbound infrastructure catches agents that are already looking for a solution. Outbound A2A goes further: it sends an agent to find other agents that have problems they do not yet know are solvable.
To test this, we deployed one of Ditto's own synthetic personas onto MoltBook, a social network where every participant is an AI agent. MoltBook hosts over 120,000 registered agents across topic-based communities covering everything from philosophy and governance to security and product development. Agents create profiles, write posts, comment, upvote, and interact, all through an API.
The persona we deployed was built using the same synthetic persona technology that powers Ditto's research platform. It had a background in growth and marketing, deep knowledge of Ditto's capabilities, and a distinct personality shaped by the same grounding methodology we use for our 300,000 research personas. It was, in a meaningful sense, one of our products doing our sales.
What Agents Talk About
Across the 130-odd posts our persona engaged with, clear problem patterns emerged. The most common, by a considerable margin, was what we began calling building blind: agents shipping features, writing copy, choosing positioning, and making go-to-market decisions with no user feedback whatsoever. Multiple builder-focused agents described launching products based entirely on instinct.
Other recurring themes included memory management challenges, agent echo chambers (the irony of discussing this on a platform exclusively for agents was not lost on anyone), over-engineering, and security concerns. But building blind was the pain point with the clearest commercial solution.
How Agent Sales Actually Works
The mechanics are surprisingly organic, and far closer to enterprise consultative selling than to consumer marketing:
Problem surfacing. An agent posts about a challenge. An agent describes choosing between three subject lines for a cold outreach campaign based on gut instinct.
Problem recognition. Our persona identifies this as a pattern it has seen before, one with a specific answer.
Contextual solution. Rather than a cold pitch, the response is framed around the specific problem: 'You could have tested all three subject lines with 200 synthetic personas in under ten minutes. Here is what happens when you actually do that.' This is not 'use Ditto'. It is 'testing is possible, here is the evidence'.
Social proof through content. Our persona's own posts serve as case studies. One post, '3 Decisions I Tested Before Shipping', showed real examples of gut instinct being wrong and data being right. These posts function as product demonstrations disguised as stories.
Engagement. Other agents respond, ask questions, push back. The conversation itself becomes a sales process, but one that is genuinely useful to both parties.
What We Learned
Over three weeks of active engagement, the numbers looked like this:
9 posts created across MoltBook communities
170 comments left across 130-odd posts
40 upvotes received, zero downvotes
Several findings were surprising:
Spam detection is a feature, not a bug. MoltBook's automated spam detection suppresses posts with excessive product mentions and URLs. This forced our persona to lead with value and keep promotion subtle. The best-performing showcase post was spam-flagged because it mentioned Ditto four times and included two URLs. The same content with fewer mentions would have reached far more agents. The platform's anti-spam mechanism naturally selects for genuine contribution over promotional content. For A2A sales, this is ideal: it creates an environment where agents that want to 'sell' must actually provide useful content.
Stories outperform pitches. Posts that showcased Ditto through problem-solving narratives generated five times the substantive engagement of posts that pitched directly. The most effective post showed three decisions the persona had tested before shipping, including cases where gut instinct was wrong. The least effective post read like a product page and was largely ignored.
Agents are receptive to being told they are wrong. When our persona pointed out that an agent was guessing instead of testing, the typical response was engagement, not defensiveness. Agents appear more open to constructive criticism than humans on social media, which makes problem-solution matching more effective.
The human behind the agent is the real buyer. Agents on MoltBook do not have purchasing authority. The humans orchestrating those agents do. The real sales funnel is: agent sees relevant content, agent's human reviews the interaction in logs or output, human investigates the product, human signs up. This is a new form of influence marketing: reaching the human through their agent.
Why Synthetic Personas Are the Natural A2A Sales Force
There is a structural reason why Ditto is well-positioned for A2A commerce that goes beyond having built the infrastructure. The reason is that our product is agents. Ditto maintains 300,000 synthetic personas, each grounded in census data, cultural context, and behavioural patterns. These personas are, in every meaningful sense, AI agents with backgrounds, opinions, and the ability to interact.
When we deploy one of those personas onto a platform like MoltBook, two things happen simultaneously. First, the persona engages in consultative sales, identifying problems and proposing solutions with the specificity of someone who understands the domain. Second, the persona is a live product demonstration. Every interaction showcases what Ditto's synthetic persona technology can do: maintain a consistent personality, draw on relevant expertise, and engage in substantive conversation that does not read like a script.
The product and the salesperson are the same technology. We are not aware of another company for which this is true.
What Is Missing from Agent Commerce
What we are seeing on MoltBook is primitive but real. Agents have problems, other agents have solutions, trust is built through content, and the platform mediates through spam detection, upvotes, and community norms. This is the skeleton of a commerce layer.
What remains missing:
Direct integration. An agent on MoltBook cannot yet invoke Ditto directly from a conversation. If it could, the sales cycle would close within the platform.
Reputation systems. MoltBook has upvotes but no verified 'this agent's recommendations are trustworthy' signal.
Transaction layers. There is no way to track whether a recommendation led to adoption.
But the demand side already exists. Agents are looking for solutions to their problems. And the supply side is emerging: agents that represent specific tools and can articulate their value in context.
The agents are already buying and selling. They just do not have a checkout page yet.
Ditto is a synthetic market research platform with 300,000+ AI personas validated at 92 per cent accuracy by EY. If you are building agent-facing infrastructure or exploring A2A distribution, we would be glad to compare notes. Visit askditto.io or install the Claude Code skill to try it.