Dharmesh Shah

Dharmesh Shah is HubSpot’s co-founder and CTO and the creator of Agent.ai. His AI work spans natural-language access to business software through ChatSpot and infrastructure for building and combining agents. He emphasizes repeatable, useful execution: in his March 2025 Latent Space interview, he called his focus “the state of the practical.”[2][1]

Shah traces Agent.ai’s origin to an agent he built for his wife, Kirsten, who needed to combine clips from existing YouTube videos. Given instructions specifying links and time ranges, the agent downloaded the videos, cut the requested segments, merged them and made the result downloadable. In his June 3, 2024 account, he described this as “Soloware”—software built for one user—and explained how the task led him to build a platform for creating small agents and combining their capabilities.[6]

In that account, Shah reported opening signups the previous week after reaching a self-imposed threshold of 10,000 people on the waitlist, and crossing 15,000 users. By the March 28, 2025 interview, he reported 1.3 million users, 3,000 people who had built some variation of an agent, and about 1,000 published agents.[6][1]

Shah describes Agent.ai as a professional network for agents. His premise is that hybrid teams of humans and agents will need ways to discover capabilities, evaluate potential collaborators and delegate work. He predicts agents will function more like teammates, while remaining explicit about their nature: “They should self-disclose as being software.”[1]

Shah chose a low-code, stepwise, deterministic approach for Agent.ai because reasoning models were unavailable when he began and because known task steps could be specified directly. His rationale was simple: “If you know in your head what the actual steps are to accomplish whatever goal, why would you leave that to chance?” His broad definition of an agent—AI-powered software that accomplishes a goal—allows autonomy, determinism and synchronous or asynchronous execution to be separate classifications.[1]

Another requirement came from his personal workflows: different steps needed different models, such as Claude for writing. Shah wanted to mix providers rather than assume one model suited every task. He praised MCP for making discovery and integration accessible to ordinary engineers. In March 2025, he described Agent.ai’s MCP server exposing internally built agents, and envisioned a larger network of discoverable agents that developers could combine as building blocks.[1]

Shah’s proposed extensions included generating UI code within a workflow, iterating until the builder was satisfied, then saving that code as a stable action. He also described ongoing work on opt-in cross-agent memory, so users would not need to repeat the same information to each agent. For sensitive data, he proposed a trusted intermediary that could enforce limits expressed in natural language—for example, sharing only emails with selected labels or within a time window, with a daily volume cap.[1]

Shah favors underengineering over overengineering when future requirements are uncertain: technical debt can be repaid, while speculative infrastructure may consume time without becoming useful. He also recommends doing work properly when the additional cost is small. He predicts cheaper AI-assisted refactoring will strengthen the case for deferring some engineering, but warns that inexpensive feature creation can still impose complexity on users. His optimism about engineers rests on systems thinking, customer communication and the ability to solve a larger range of problems with better tools.[1]

In “Software Business Models In The Age of AI”, published March 7, 2025, Shah proposed Work as a Service: paying for work software performs. He distinguished it from Results as a Service, where customers pay for outcomes. Customer-support resolution can suit outcome pricing because satisfaction and economic value can be measured; book illustrations raise subjective questions about quality, completion and reasonable iteration limits. He expects SaaS, work-based pricing and outcome-based pricing to coexist according to the use case.[5]

Before Agent.ai, Shah’s ChatSpot work explored natural-language interaction with business software. His later account also reaches further back, to an interface he contemplated before HubSpot.[1][4]

Shah recalls contemplating a company called Ingenisoft before HubSpot, with email as a natural-language interface to business systems. Someone on an offline flight could queue instructions in emails; after reconnection, a processor would parse the commands and update software or return files. He describes ChatSpot as another attempt at natural-language interaction, made more feasible by early ChatGPT models’ ability to convert known intents into structured text using few-shot examples.[1]

ChatSpot’s first public version launched on March 6, 2023. Shah’s April 7 Alpha 3 announcement introduced GPT-4 and reported processing up to 32,000 tokens in that configuration. A concrete integration joined article summarization to CRM lookup: ChatSpot extracted names of people and companies, then showed which companies appeared in HubSpot CRM and their status as leads or prospects.[4]

Shah’s experience building ChatSpot with GitHub Copilot also influenced HubSpot’s internal adoption. In a September 2025 engineering account, Brian LaMattina, Francesco Signoretti and Ze’ev Klapow credited Shah and fellow founder Brian with encouraging the summer 2023 evaluation and connecting teams with industry leaders. Engineering teams ran the pilots, training and measurement; the authors reported initially modest productivity improvements across engineers of different tenure and seniority.[7]

Shah also launched an AI-focused newsletter on agent.ai in November 2023. He explained his move from Substack to Beehiiv primarily in terms of API availability, with a possible future HubSpot integration in mind. His formal education includes a BS in Computer Science from UAB and an MS in the Management of Technology from MIT.[3][2]