There's a particular kind of panic that sets in around day three of an AI agent deployment. The coding assistant that confidently closed GitHub issues on Tuesday has, by Thursday morning, quietly derailed. It's reassigned the wrong Linear tickets, flooded your Slack channels with nonsequiturs, and triggered a cascade of false alerts in Sentry. The CEO wants to know what happened. Your engineers are manually rolling back changes. And somewhere in a Slack thread marked "urgent," someone types the question everyone's thinking: "Can we just turn this thing off?"
This isn't edge-case paranoia. It's the defining constraint that will determine which companies actually capture value in what analysts project as a $183 billion AI agent market by 2033—and which ones join the 40% of agentic AI initiatives Gartner now expects to be canceled by 2027.
Enter Polymath Labs, a two-person team emerging from Y Combinator's Winter 2026 batch with an unusual pitch. They're not building another coding agent. They're not promising to make agents smarter, faster, or cheaper. Instead, they're constructing what they call "training grounds" for long-horizon autonomy—production-grade reinforcement learning environments where AI agents can be trained and stress-tested across days and weeks of multi-tool workflows, not just the sanitized minutes most benchmarks measure.
Whether that matters—whether it's infrastructure the industry desperately needs or an elegant solution to a problem the market will solve differently—is about to become clear.
The Appetite Is Real. The Execution Isn't.
Enterprise interest in AI agents has moved beyond pilot purgatory, at least on paper. McKinsey's 2025 State of AI survey found 62% of companies actively experimenting with agents, while Gartner predicts that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from under 5% in 2025. The infrastructure layer is maturing rapidly. AWS launched multi-agent collaboration on Bedrock this year. Microsoft shipped Agent 365, positioning AI agents as managed digital workers. The Agentic AI Foundation—backed by OpenAI, Anthropic, and Block—is working to standardize protocols like Model Context Protocol across the ecosystem.
Market forecasters are bullish, perhaps aggressively so. Multiple analyst firms peg the global AI agents market at roughly $7.6 to $8 billion in 2025, growing to $183 to $251 billion by 2033 or 2034. That implies a compound annual growth rate approaching 46% to 50%. The reinforcement learning market, which underpins much of this agent training, is tracking from $12.4 billion this year to $111 billion by 2033, according to industry projections.
But there's a widening gap between the projections and what actually survives contact with production.
When the Leaderboard Meets Reality
The numbers tell the story more clearly than any enterprise case study. On SWE-bench Verified, a curated collection of GitHub issues, leading models from Anthropic, OpenAI, and Google routinely post success rates between 70% and 80%. Switch to SWE-bench Pro—a harder, less contaminated evaluation—and those same models plummet to around 20% to 23%.
Newer long-horizon benchmarks paint an even starker picture. SWE-EVO, which tracks repository-level changes averaging 21 files per task, sees GPT-5 with OpenHands scoring roughly 21%. LongCLI-Bench, focused on command-line workflows that unfold over time, reports state-of-the-art performance below 20%.
The pattern holds across domains. Short, tightly bounded tasks? They work. Anything requiring sustained tool orchestration, multi-day context retention, or graceful recovery from partial failures? Not so much.
ServiceNow has publicly claimed 410,000 hours saved annually and $17.7 million in cost avoidance from internal agent deployments. But those figures come from highly controlled, purpose-built environments—bespoke systems designed around known constraints. The median company attempting to replicate that success encounters what Gartner delicately terms "unclear value" and "inadequate risk controls." Translation: agents that quietly drift off-task or catastrophically misinterpret instructions over longer time horizons.
Building the Infrastructure Beneath the Agent
Dylan Ma, previously at Hume AI and AWS, and Naren Yenuganti, who spent time at Plaid and Amazon, are targeting the substrate layer beneath the agent itself. Their thesis, refined through months of customer conversations and technical prototyping, comes down to this: you can't train reliable long-horizon agents without environments that mirror the messy, multi-system reality of how work actually happens.
Polymath's product centers on what they call "world generation models"—systems that programmatically create reinforcement learning training environments with five core components. Applications and tools: Slack, Linear, Sentry, GitHub, CI/CD pipelines, cloud monitoring dashboards. Data: realistic state snapshots and interaction histories. Tasks: progressively complex workflows that require coordination across systems. Verifiers: automated checks for successful outcomes, not just completion signals. And the agent itself, operating across all of it.
The goal is to automate the construction of end-to-end scenarios that span days or weeks. Most existing benchmarks measure minutes, maybe hours. Real workflows don't operate on that timeline.
They've started with software engineering, but deliberately outside the IDE. Their first public benchmark, Horizon-SWE, orchestrates live applications across Slack notifications, Linear ticket management, Sentry error tracking, Prometheus metrics, and CI/CD systems—the operational context where most engineering work actually unfolds, not just where code gets written. Early results, posted publicly and highlighted by Y Combinator, show leading models landing somewhere in the low-20s to roughly 25% success range.
That's not impressive in absolute terms. But it's a useful baseline for the kind of end-to-end autonomy enterprises say they need—and currently can't get.
The pitch to frontier labs and enterprises is straightforward. If you're building or deploying agents that need to operate autonomously across complex, multi-day workflows, you need realistic environments to train and validate them before they touch production. Static benchmarks measure model capability at a point in time. Dynamic RL environments surface the credit assignment problems, planning failures, and tool-chaining errors that don't show up until agents run for days without human intervention.
Polymath isn't alone in this. Recent research frameworks like LOOP (an efficient PPO variant for interactive agents) and WebAgent-R1 (multi-turn RL on web tasks) demonstrate that reinforcement learning can meaningfully improve how agents recover from failures and use documentation. Process-level reward systems like RLVMR have shown state-of-the-art results on long-horizon reasoning benchmarks by decomposing tasks into verifiable intermediate steps. The academic community seems to be converging: long-horizon reliability requires dense, structured feedback loops that supervised fine-tuning alone can't deliver.
The risk surface is also expanding in ways most enterprises haven't anticipated. AgentLAB, a new benchmark for long-horizon attacks, documents vulnerabilities unique to autonomous agents—intent hijacking, task injection, memory poisoning—that single-turn defenses fail to address. As agents operate autonomously over longer timescales, the window for adversarial exploitation grows. So do the stakes.
A Crowded Field with Unclear Winners
The competitive landscape is crowded but oddly segmented. OpenAI launched Operator, an RL-trained agent for autonomous web browsing and computer use, alongside Deep Research for multi-day research workflows. Anthropic's Claude models dominate many SWE-bench leaderboards and pioneered computer use capabilities. Google has invested heavily in environments like AppWorld for desktop and web evaluation. UiPath claimed the top spot on OSWorld-Verified in January 2026 with their Screen Agent. Simular reported surpassing human baselines in December 2025.
But most of these are agent products or model providers competing at the application layer. Polymath is positioning as infrastructure—the layer that trains and evaluates agents, not the agent itself. Their implied customer profile suggests frontier labs needing realistic post-training environments and enterprises requiring verifiable, end-to-end validation before they commit budgets to autonomous workflows. The Y Combinator network provides initial go-to-market leverage, though they haven't publicly named design partners or customers as of late February 2026.
The strategic question hanging over all of this: does the training and evaluation substrate become a standalone category worth defending, or does it get absorbed into model providers and cloud platforms with distribution advantages Polymath can't match? AWS, Microsoft, and Google are all building multi-agent orchestration directly into their clouds. Cognition Labs' Devin has raised significant capital selling an end-to-end software engineering agent. GitHub now offers Agent HQ to manage multiple coding agents. If Polymath's world generation models prove essential to the stack, the logical acquirers are already circling.
The Next 12 Months Will Tell
Whether long-horizon reliability is solvable with better training infrastructure—or fundamentally limited by current architectures—should become clearer over the next year to 18 months. Several indicators matter more than others.
Does Horizon-SWE gain traction as an industry-standard benchmark, or does it remain a niche evaluation? Do frontier labs adopt world generation approaches in their reinforcement learning pipelines, or do they build proprietary alternatives? And most importantly: do enterprises see measurable improvement in agent stability that translates to production deployments at scale?
Standardization momentum is building, which could work in Polymath's favor. The Agentic AI Foundation is working to unify agent interoperability standards. Secure variants of Model Context Protocol are in active research. The EU AI Act's high-risk obligations kick in fully by August 2026, pushing companies toward verifiable evaluation and logging—precisely the kind of infrastructure Polymath is building. NIST's AI Risk Management Framework, published in July 2024, provides implementation guidance for trustworthy AI that aligns closely with robust agent evaluation.
The market projections, ambitious as they are, assume reliability improves fast enough to justify continued enterprise investment. Mayfield's 2026 CXO Network survey found 42% of companies already have agents in production, with line-of-business leaders increasingly driving purchasing decisions rather than waiting for IT approval. But Gartner's dual prediction—40% adoption and 40% cancellation rates—suggests the market will bifurcate sharply. Companies that solve reliability will scale. Those that don't will retrench.
For Polymath, the bet is that the training substrate becomes critical infrastructure as the industry shifts from Q&A copilots to autonomous execution. Whether that materializes as a venture-scale outcome, an acquisition by a larger platform, or an open standard that others adopt depends on a question no benchmark can answer yet.
Can two founders and a batch of Y Combinator funding actually close the reliability gap? Or is the gap wider—and the solution more capital-intensive—than anyone currently positioned in this market wants to admit?
The agents are running. The clock is ticking. And somewhere between hour six and day three, we're about to find out.
