Bespoke Labs Raises $40M to Build the Training Worlds AI Agents Actually Need

Bespoke Labs just closed $40 million in combined seed and Series A funding to do something most AI companies treat as an afterthought: build the environments that actually train reliable agents. The company's thesis is deceptively simple. Compute is cheap, base models keep improving, but the one thing that will determine whether an agent can be trusted in production is the quality of the world it learned in.

The funding, broken down

Bespoke Labs raised the bulk of the funds, $31.75 million, through a Series A round led by Wing VC. The company earlier raised $8.25 million from a consortium that included Google DeepMind chief scientist Jeff Dean. The investor list reads like a who's who of AI infrastructure bets:

• Series A: Wing VC led, with participation from Mayfield, The House Fund, dbt Labs CEO Tristan Handy, and angel investors from Anthropic, OpenAI, and Meta.
• Seed: Led by 8VC, with participation from Jeff Dean, Resolve AI CEO Spiros Xanthos, and DevRev CEO Dheeraj Pandey.

Bespoke Labs plans to use the capital to expand its research team, scale its environment-building infrastructure, and accelerate its business momentum. The simultaneous announcement of both rounds is a signal that the company was heads-down building before taking a victory lap.

Why agents keep failing

Today's AI agents are powerful but unreliable. They can write code, answer questions, and complete short tasks, but they still struggle to operate autonomously over hours or days the way a human coworker does. This is not primarily a model problem. Post-training -- the phase where a model is fine-tuned and shaped for specific behaviors after its initial large-scale training -- is where reliability is won or lost. And post-training requires high-quality environments to train in.

Independent benchmarks from METR show that the length of tasks AI agents can reliably complete has been doubling roughly every seven months. Sustaining that trajectory requires environments that grow in complexity at the same pace, which is exactly the problem Bespoke Labs was built to solve. The implication: if you extrapolate that curve to the end of the decade, you need agents that can stay coherent across multi-day workflows. The environments to train those agents don't exist yet.

What Bespoke actually builds

Bespoke Labs is behind the environments and tooling that sit underneath reliable agents. The company designs and builds environments that look and behave like real companies: large codebases, microservices, realistic logs, tickets, email, Slack, and more, so agents can learn long-horizon workflows that are economically meaningful. Think of it as a flight simulator for AI agents -- except the simulator has to be indistinguishable from a real company.

The platform has three core layers:

• Environment engine: Lets expert networks and automated systems compose realistic, multi-tool, multi-step worlds far faster than hand-coding each one.
• Sandboxing and execution layer: Runs the AI environments using a sandboxing layer that helps minimize latency and boost throughput.
• Agent optimization layer: Uses techniques like GEPA and reinforcement learning on top of those environments to improve a customer's agent.

GEPA: the optimizer that beats RL with fewer shots

One of Bespoke's most technically interesting contributions is GEPA (Genetic-Pareto Agent Optimizer), an open-source agent optimizer that takes a fundamentally different approach to improving agent behavior. Instead of relying on thousands of gradient-descent steps, it uses evolutionary search and natural language reflection.

Given any AI system containing one or more LLM prompts, GEPA samples trajectories and reflects on them in natural language to diagnose problems, propose and test prompt updates, and combine complementary lessons from the Pareto frontier of its own attempts. As a result, it can often turn even just a few rollouts into a large quality gain. Across six tasks, GEPA outperforms GRPO by 6 percentage points on average and by up to 19pp, while using up to 35x fewer rollouts.

The "Pareto" part is key. The core insight is that different prompts may excel at different types of problems. GEPA addresses this by maintaining a diverse pool of candidate prompts rather than converging on a single "best" one. It's the difference between a specialist and a well-rounded agent. Shopify CEO Tobi Lutke has already called it severely underrated in the AI engineering world.

# Install GEPA with LangChain support
pip install "gepa[langchain]"
# Or with DSPy
pip install "gepa[dspy]"
# GEPA connects to your system via the GEPAAdapter interface
# Implement evaluate() and make_reflective_dataset(), GEPA handles the rest

The open-source track record

Bespoke's credibility isn't just the pitch deck -- it's the shipped work. Bespoke Labs is a core contributor to Terminal-Bench, one of the most widely cited benchmarks for agentic capability, and the team behind OpenThoughts, an open reasoning dataset that has been downloaded more than 500,000 times and used by groups including Thinking Machines Lab, Meta, and Amazon.

This open-research approach is also a competitive moat. The team is made up of research scientists and engineers, not the project managers and contractors that other companies rely on. Many startups in the agent space rely on manual contractors and basic application-level adjustments to tweak performance. Bespoke Labs takes a strictly research-first approach. That distinction matters: the quality of an RL environment is only as good as the research behind it.

Who wins, who should pay attention

The clearest winners are frontier labs and enterprises already deploying agents. "As frontier labs and AI-native enterprises push the boundaries of long horizon agentic capabilities, a new generation of data and training infrastructure is required," said Peter Wagner, founding partner of Wing Venture Capital. Anthropic, OpenAI, and Google DeepMind already use Terminal-Bench to showcase their models' agentic abilities -- meaning Bespoke is already embedded in the evaluation pipelines of the biggest labs.

The competitive pressure falls on Scale AI and similar data infrastructure companies. Scale AI provides the data infrastructure and reinforcement learning pipelines that power AI agent training for government, defense, and enterprise clients -- a direct overlap with Bespoke's ambitions. The difference is Bespoke's research-first DNA and its focus specifically on RL environments rather than broad data labeling.

The hard problem ahead

The funding unlocks a genuinely unsolved research agenda. Bespoke's blog lays out the questions keeping the team up at night:

• How do you measure environment quality before spending the compute to find out?
• How do you build worlds complex enough for agents that need to stay coherent across multi-day workflows?
• How do you generate realistic environments at scale using AI-driven pipelines instead of hand-built apps?
• How do you simulate an entire company -- Slack, email, Jira, microservices, other agents -- as a training ground?

That last question is the frontier. The same research muscle powers the company's environment work: novel curation pipelines, methods for generating environments synthetically, and techniques for snapshotting real-world infrastructure and recreating it as a training environment. If Bespoke can crack synthetic environment generation at scale, it changes what's possible for every organization trying to build production-grade agents.

For teams actively building or evaluating agents today, GEPA and OpenThoughts are worth examining immediately -- both are open-source and already in use at major labs. The broader platform is what the $40M is building toward, and the official announcement signals Bespoke is ready to move from research lab to infrastructure company.