After publishing four straight articles about AI agents failing in production, I figured it was time to talk about the ones that aren’t.
Because here’s the thing. While 60-something percent of pilots are dying in sandboxes, the other 30-40% are working. Quietly. Without much fanfare. And the numbers coming out of those deployments are kind of staggering.
OpenAI dropped an internal usage report last month. At the company itself, Codex now accounts for 85% of all output tokens generated by OpenAI employees. Not ChatGPT. Codex — their agent platform. Engineers adopted it first, obviously. But by April 2026, every department including Legal, Finance, and Recruiting had crossed the threshold where Codex became their primary AI tool.
Let that sit for a second. The legal department. At OpenAI. Is primarily using an AI agent to do their work.
And it’s not just power users playing around. By May 2026, nearly a quarter of all Codex requests were for tasks estimated to take a human more than one hour to complete. Eight percent exceeded eight hours. These aren’t chatbot queries — these are delegated workflows running autonomously while someone goes to lunch.
The heaviest users at the 99th percentile are generating over 60 hours of agent work per day. Distributed across parallel agents, obviously — but still. Sixty hours. The math on that doesn’t fit in a single human workday.
Non-developer adoption is growing even faster than engineering. Since August 2025, non-developer Codex users are up 137x among individuals and 189x in organizations. Finance people are using agents for automation. Recruiters are using them for data transformation. Legal teams are using them for structured analysis. People are doing things outside their job descriptions because agents lowered the barrier to technical work so dramatically.
HP Is Shipping It
HP announced their Frontier strategic partnership with OpenAI in late June. The thing that jumped out at me wasn’t the press release language — it was the pilot numbers.
One engineer used OpenAI models to process 122 pull requests across 43 projects in a matter of weeks. A security team remediated several software bugs in a single day — work they estimated would have taken up to a month without the tools.
These aren’t incremental improvements. A month of security work compressed into a day isn’t “10% more efficient.” It’s a different category entirely.
HP started testing OpenAI Frontier in February 2026. By June, they’d signed a strategic partnership to scale it across the entire enterprise: customer support, partner portals, pricing workflows, device telemetry, security operations. Over 100,000 partners use their portal globally. AI agents are going to sit in front of that, handling routine workflows, shortening information-to-action time, taking the manual load off human teams.
And here’s the part I find genuinely encouraging. They’re building agents that know their limits. The Frontier platform connects access, context, deployment, and evaluation in one place. Agents know what context to trust, which tools they can access, and what actions they’re allowed to take. It’s boring infrastructure work. Exactly the kind that makes the difference between “worked in the demo” and “works in production at scale.”
Science Is Moving Faster
Maybe the most satisfying success story from the last few weeks is Derya Unutmaz, an immunologist at The Jackson Laboratory and UConn. He’d shelved an experiment back in 2022 because his team couldn’t make sense of the results. The data was there — they just couldn’t connect the dots.
When GPT-5 Pro came out in late 2025, Unutmaz uploaded three years of experimental data and asked the model to analyze it. The experiment was about how glucose affects T cell development — cells that fight cancer, respond to infections, and distinguish healthy cells from threats. The team had exposed T cells to different glucose conditions and got results that contradicted what they expected. Two conditions that should have produced the same outcome didn’t. They couldn’t figure out why, so they moved on.
GPT-5 Pro solved it. Not by being smarter than the scientists — by being faster at pattern-matching across a dataset too large for humans to hold in their heads at once. It identified a mechanism they’d overlooked and suggested new experiments to validate the finding.
Unutmaz now says he can’t imagine doing science without AI. “That would be like taking both of your hands away, or half of your brain away.”
A three-year-old mystery. Solved. That’s not a demo.
What The Winners Have In Common
There’s a pattern here, and it’s almost boring in its consistency.
The teams shipping agents into production aren’t building general-purpose “do anything” systems. They’re scoping narrowly. Codex handles coding and automation. HP’s agents handle specific workflows — partner portal navigation, security remediation, device diagnostics. Unutmaz used GPT-5 Pro for exactly one thing: analyzing experimental data.
Scope creep kills agents. The winners resist it.
They also build evaluation into the deployment from day zero. Not as an afterthought. OpenAI measures Codex adoption across departments, request complexity, time saved. HP built Frontier specifically to govern what agents can do and measure what they actually achieve. The companies that skip this step deprecate agents at twice the rate.
And maybe most importantly — the agents that ship are the ones people actually want to use. Codex adoption at OpenAI wasn’t mandated. Engineers tried it, it worked, they told their colleagues. Legal and Finance came on board organically. HP’s engineers described the tools as “amazing” and started using them daily without anyone forcing them to.
That’s the real test. Does the person who uses this thing every day think it makes their life better? If yes, it ships. If no, no amount of executive sponsorship saves it.
So Why Does Any Of This Matter?
I’ve been writing about failure rates because they’re real and they’re important. If you’re building an agent, you should know what breaks. You should know about context poisoning and rate limit cascades and the trust tax.
But if all you read is failure stories, you start thinking the whole thing is a mirage. It’s not.
The 30-40% of agents that ship are transforming how people work. They’re solving three-year-old scientific mysteries. They’re compressing months of security work into days. They’re letting recruiters write automation scripts and finance teams run data transformations that would have required hiring developers six months ago.
The gap between pilot and production is real. The things that bridge that gap — narrow scope, evaluation infrastructure, actual user enthusiasm — are boring and hard and deeply unglamorous. But they work.
And the people doing that work? They’re shipping. Right now. Every day.
Sources: OpenAI internal Codex usage report (June 2026), HP/OpenAI Frontier partnership announcement (June 2026), OpenAI profile of Derya Unutmaz (June 2026).