Bespoke Labs secures $40 million to develop dependable AI agents.

      AI agents are capable of writing code and answering queries, but they struggle with lengthy and complex tasks. A startup from Mountain View has just secured $40 million to create training environments that aim to address this issue. Bespoke Labs, focused on developing the environments for training and testing AI agents, announced the funding, which includes a Series A led by Wing VC and a previous seed round led by 8VC. Notable investors include angels from Anthropic, OpenAI, and Meta, alongside Google DeepMind's Jeff Dean and dbt Labs CEO Tristan Handy.

      Training grounds for agents

      Current agents are proficient but inconsistent—they manage short tasks effectively but often fail to operate independently over extended periods, as a human colleague would. Bespoke believes that the solution lies not in creating larger models, but in enhancing the training environments.

      To this end, Bespoke creates simulated versions of actual companies, incorporating elements like extensive codebases, microservices, logs, support tickets, emails, and Slack conversations. Agents learn in these simulated contexts, mastering the long, multi-step workflows necessary for effective work. Bespoke also assists clients in assessing and optimizing their agents through an in-house optimizer called GEPA, which identifies improved prompts and strategies more swiftly than manual tuning.

      Research lab, not a contracting firm

      Established in 2024 by CEO Mahesh Sathiamoorthy and chief scientist Alex Dimakis, Bespoke's approximately 40-member team is research-oriented. The company is a key contributor to Terminal-Bench, a well-known test of agent capabilities, and developed OpenThoughts, an open reasoning dataset that has been downloaded over 500,000 times by organizations, including Meta and Amazon.

      Rather than outsource tasks to contractors, Bespoke views the creation of environments as a research endeavor and commercially offers the resulting infrastructure.

      Significance

      Bespoke strategically timed its launch. Independent assessments by METR indicate that the duration of tasks agents can reliably complete has been doubling roughly every seven months, with some analyses noting it could be as short as four months. Maintaining this trend requires environments that also increase in complexity at a similar rate, which is exactly what Bespoke offers.

      The competitive landscape is dense, with many companies addressing agent reliability through various approaches, including self-learning systems and businesses focused on stress-testing, evaluating, and benchmarking agents prior to release. Others are exploring the cost-efficiency of operating agents at scale. Bespoke is betting that the training environment, rather than the model itself, will determine which agents successfully transition to production.

      The question of whether superior environments can outperform larger models remains unanswered. The response will influence which companies endure through the upcoming funding rounds.