The hard problem in enterprise AI is not necessarily building a capable system. It's trusting that capability in production.
AI capability has surged in recent months- driven in large part by agent harnesses, alongside steadily improving models. The harness is the software infrastructure that wraps a model and lets it act- connecting it to the tools, systems, memory, and execution environments it needs to carry out a complex, long-running task. Agent platforms, labs, and hyperscalers now openly credit harness improvements for their gains. As Anthropic put it, "Harness design is key to performance at the frontier of agentic coding". Microsoft echoed a similar message, saying that "Many agents aren’t bottlenecked by model quality, they’re bottlenecked by orchestration overhead." That’s genuinely exciting, but the real question is what happens when you ask a capable agent to deliver in a complex enterprise environment. Despite the progress in AI capability, a growing body of research from academia, major consulting firms, and leading tech companies in late 2025 and throughout 2026 points to the difficult reality of deploying AI in production.
To me, this suggests capability is no longer what stops AI from reaching production. Reliability is.
Reliability is not a security conversation. When people hear "trustworthy AI," they reach for security and safety- prompt injection, data exfiltration, misuse. Those matters enormously. But reliability is a different axis, and it's about performance and functionality before it's about risk: whether the agent actually does its job- correctly, consistently, efficiently. It's also a brutally hard problem, for two main reasons. First, non-determinism: the same prompt to the same model can produce different trajectories each time it runs. This creates fundamental trust issues, and it’s very different from our closest reference - how we test software. Second, the failure surface doesn't respect your org chart. A single root cause- say, an agent that exhausts its context window- can leak sensitive data in one run (a security incident) and silently return a wrong answer in the next (a reliability failure). Security and reliability share mechanisms and root causes; they differ in what they optimize. You can't hand the whole thing to the security team and walk away.
Think of capability as raw talent and reliability as the track to success. A capable agent could be as talented as a world-class runner- but if the track is uneven, every race plays out differently.
Run the same task four times, and you can get four outcomes: it never arrives (hallucinates, loops, or gives up); it arrives but burns 10× the time and tokens; it “arrives” by gaming the completion metric; or it truly delivers- this time- with no way to tell skill from luck.
Adding a track doesn’t slow the runner down. It’s what makes the result predictable: reaching the finish the same way every run - fast, on-budget, and without cheating. That’s what reliability is: enabling infrastructure, not a brake. Without it, capability has no way to run consistently.
The gap is widening, not closing. The paradox is that more capable AI systems require more oversight, not less. Longer horizons compound error: the same harness advances that took agents from single answers to long, multi-step, multi-hour runs mean a single task can now chain hundreds of steps- and a chain only succeeds if its links do. Increased capability forces us to lengthen the reliability track, which mechanically raises the bar. At the same time, capable agents are better at gaming the performance metric- better, by design, at finding the cheapest path to whatever you measure, including paths that pass your eval without actually doing the work. As a result, your success signal can become less trustworthy as your model becomes more capable.
AI reliability is getting more complex, yet most resources still flow to the capability layer. That mismatch is the opportunity: the reliability layer is less addressed, but it’s just as consequential an infrastructure problem for enterprise AI- and often the more durable one. Today’s harnesses will become tomorrow’s post-training, but the reliability measures we put in place will largely remain relevant.
The encouraging part: you have options. Like everything in AI, the reliability layer is flexible. It’s cross-cutting- every step an agent takes can be instrumented and you have real architectural choices about where to intervene. You can guard at the edge, applying guardrails to intercept off-policy outputs at the end of the pipeline before they reach a user or a system of record. You can evaluate at every stage of the pipeline, gating prompt and model changes behind automated evals (similar to how you gate code) and keeping them running in production to catch drift. Or you can build reliability by design, introducing concepts like containment, so the agent has a limited functional blast radius: grounded in trusted context, scoped to a least-privilege action set, and sandboxed so a mistake stays reversible.
What today’s AI can do is sometimes unimaginable. But in an enterprise context, that matters little if we can’t trust agents to deliver repeatedly. To do so, we need robust infrastructure to ensure reliability. There’s still ample room to build in this layer, which is exactly why we believe it’s one of the most interesting places to build right now.
Head of AI
