What happened
Every employee is put on an AI coding agent. Individual output jumps. Within weeks the work is stuck in the same four places it was always stuck: review, deploy, data access, validation. Everybody is ready and nobody can say yes.
The agentic OS goes live. Identity per agent, brokered credentials, adversarial review in the build path, containment at the compute boundary, and signed evidence for every action.
486 apps and 42 agents, 20,372 deployments, 30.8 million requests. Finance, HR, marketing, revenue operations and product all shipping into production.
The work stopped queueing
Ask what a platform like this is worth and you get a productivity percentage back. The value showed up somewhere else: in things that exist now and otherwise would not, because the queue to build them used to outlast the idea. A new foundation model reached a customer's on-premise environment in 48 hours, against a 3 to 4 week estimate. Marketing built its own account-based platform end to end. One product manager designed, built and launched the Nexar Risk Index on their own. An agent now produces the annual compliance package and answers ad-hoc regulator requests.
There are also 42 agents, built by 12 people. They run on their own, triggered by a schedule or an event instead of a person opening a page, and 198 of their tasks fired straight from GitHub. The platform never instrumented them individually, so they appear in none of the explorers below and their individual impact is unmeasured.
Usage compounded while building slowed
Apps created per month fell from 150 in January to 24 by June. Requests per month went the other way, 3.3 million to 6.2 million. The apps built early get used more every month. That gap is the difference between a hackathon and an operating system.
A few of them turned into load-bearing infrastructure. The busiest one served 4.2 million requests across 138 active days.
The bill follows the use
The median app costs nothing. Not a rounding error, literally $0.00: 374 of the 486 come in under a cent a month, because an app nobody is using consumes nothing but a row in a registry.
That is what makes it safe to let everyone try. An idea nobody ends up needing never draws on anything. Failure is priced the same way: of 20,372 deployments 86.9% succeeded and the rest cost a retry, and 64 apps have been deleted outright without leaving a bill behind.
What the bill does track is the work, and the work has nothing to do with how many people are involved. The most expensive app in the estate cost $915 last month and has three users. It takes dashcam video and runs a collision-prediction model over it, so a single request is minutes of compute. Three researchers get a training and inference tool that used to be a manual pipeline, for the price of a rounding error on one salary.
The inverse holds too. The dashboard serving 4.2 million requests costs about $4 a month, because each of those requests is trivial. Nobody sized either of them. The same platform priced a heavy tool for three people and a light one for the whole company correctly, without anyone capacity-planning per app.
What people built when the queue disappeared
Eight production apps. Eight different builders. Browse the work, who owned it, and what changed for the business.
One product manager launched a revenue-generating analytics product.
Insurance and autonomous-vehicle companies wanted road-segment risk data across America, so Nexar turned it into a subscription analytics product.
Every app, every day, for 230 days
All 457 apps as one line each. Drag the cohort slider from the whole estate down to the single busiest app and watch how concentrated real usage is. The top five carry 64% of all traffic. The long tail costs almost nothing and is still worth keeping.
Open the usage explorer →What the bill actually follows
Every app plotted by what it costs against how many people use it. The 374 that cost nothing sit on the floor. The expensive ones are up and to the left, serving a handful of people each, because they do heavy work per request. Flip to the bracket view and the dots re-form into price bands: 374 apps under a cent, 57 of them at exactly nothing, against 29 apps above $100 that carry $7,924 of the bill.
Open the cost explorer →It spread like a contagion
This spread by use. Somebody opened an app a colleague had built, understood in that moment that they could do the same thing themselves, and did. There was no enablement programme and no phased plan, which is why it looks nothing like the adoption curves companies usually publish. Every action was instrumented, so we can show the mechanism working instead of asserting it.
- 76 of 77 builders used a colleague's app before building their own. One exception in the entire company. Use came first, every time.
- One new builder per existing builder. The rate lands on exactly 1.00, the threshold at which a contagion sustains itself, but it is carried by 17 people. Sixty of the 77 converted nobody.
- Three apps seeded roughly half the company. A leaderboard, a hello world and a ticket helper. Which app someone met first decided whether they ever built: one converted 26 of the 41 people it reached, another converted none of four.
- It saturated rather than faded. New builders per month fell from 47 to 1 because it ran out of people left to convert, not because interest died.
Thirteen ways to look at it
Each opens on its own. If you open only two, take what actually converts and one person's path. The four timeline views share a scrubber, so you can play the spread day by day from December to July, and the whole set is also browsable as one gallery.
The guardrails are what made the freedom possible
None of this is a story about loosening controls. It is the opposite. Letting 77 people deploy into production is only a defensible decision if something in the path can refuse, and can still refuse late enough to matter.
So the platform sat in the critical path of every action. Adversarial review ran in the build path before anything was promoted. Enforcement lived at the compute boundary, where a running process can actually be halted while it is still running. Every action landed in signed evidence the platform itself cannot rewrite.
Over seven months that produced 6,974 security reviews and 1,042 publish gates, sent 33 apps back for fixes, rejected 5, and quarantined 258 times across 18 apps. Among the findings were 1,143 prompt-injection detections in production.
A control that can only alert has to ration who builds, because every alert lands on a human queue and the queue is the chokepoint. A substrate that can contain can afford to say yes to everyone. That is the trade. The freedom and the enforcement are the same decision.
The onramp mattered as much as the guardrail
A financial controller does not want a credential. They want their app to read the warehouse, post into Slack, call an internal service and run on a schedule, without owning the blast radius when any of it goes wrong.
So the platform brokered all of it. Each of the 486 apps got an identity of its own instead of borrowing a person's. Secrets went into 181 apps at the moment of use and were never handed over, so no credential ever sat in a session or a repository. Reaching a system of record meant filing a request and getting a scoped grant: 553 access requests and 49 warehouse requests, against a catalogue of 6,113 internal endpoints an app could be granted access to.
That is the part that let people outside engineering build on the systems the company actually runs on. Fifty-five of these apps run on schedulers and 25 are wired into Slack, built by people who never had production access and never needed it.