Roughly 99% of companies say they plan to put AI agents into production. About 9 to 14% have actually done it. That gap should sober up every agent roadmap, and this month it got a name, courtesy of a report from Ness Digital Engineering published around 19 August 2026. They call it the Death Valley, the stretch between a proof of concept that dazzles in a demo and a system anyone trusts to run unattended.
The framing is dramatic, but the data behind it keeps turning up from different directions. Separate research this year found enterprises running around a dozen agents on average, with roughly half of them working alone rather than coordinating. Another read put the share of agents reaching production at about 11%. Whichever number you like, the shape is the same: everyone is piloting, almost nobody has agents actually running in production.
Why pilots die in the valley
The seductive part of an agent demo is also the reason it stalls. A scripted demo runs one happy path in a clean environment. Production is thousands of messy paths, stale data, permission edge cases and the occasional confidently wrong answer that costs real money. Ness pins the gap on trust in probabilistic systems, and that is fair. A tool that is right 92% of the time is a miracle in a demo and a liability in a workflow where the other 8% issues refunds nobody signed off on.
The other killers are less philosophical and more boring. Teams jump into pilots with no production-readiness criteria, no observability, no plan for who owns the agent when it misbehaves. A survey of teams running agents in production found reliability, not raw model capability, was the top development challenge, and most cap their agents at around ten steps before a human has to step in. That cap is a confession: nobody trusts the thing to run long without supervision.
What separates the 10% that make it
The teams that cross the valley treat the agent as a system, not a chatbot with ambitions. In practice that means:
- Least privilege from day one. Inventory every tool and system the agent can touch, and give it the narrowest access that lets it do the job.
- Observability that logs every tool call. If you cannot replay what the agent did, you cannot debug it, audit it, or defend it to compliance.
- Human approval on the actions that hurt. Payments, code merges, data exports. Cheap reversible actions can run free; expensive irreversible ones need a gate.
- Production criteria before the pilot, not after. Decide what ‘good enough to trust’ looks like in numbers up front, then hold the pilot to it.
The unfashionable takeaway
Most agent pilots deserve to die in the valley. They were built to impress a steering committee, not to survive a Tuesday. The useful reframe is to stop treating the demo as the hard part. Getting an agent to look brilliant once is easy now. Getting it to be reliable, observable and boring enough to trust every day is the actual work, and it is where the 90% fall down. Budget for that, or enjoy your very expensive proof of concept.
Did you know: the ‘Death Valley’ tag is borrowed from startup finance, where it describes the period after launch when cash is burning and revenue has not arrived. Agents have found their own version, minus the venture funding.
Sources
- ANI: 99% plan agentic AI, only 9-14% in production
- Teradata: why agentic AI stalls
- Belitsoft/Barchart: enterprises run 12 agents, half work alone
Related on Top Tool Stack: Claude Code, Codex or Cursor: Which Agent Harness Wins Now · Google Moves Its A2A Agent Protocol In With MCP