In 1945, as postwar governments sought to reshape the world order, Friedrich Hayek argued that central planning would fail, despite the planners’ intellect and competence, because the knowledge an economy runs on “never exists in concentrated or integrated form, but solely as the dispersed bits of incomplete and frequently contradictory knowledge which all the separate individuals possess.”

He may as well have been speaking about the industrial supply chain, and the intricacies of the labor and staffing of these supply chain businesses. The Traba operator managing a customer’s distribution center in Detroit knows a worker running five minutes late has missed the shift, because the crew drives off without them. Another knows a site can’t take same-day backfills, because its process requires workers to bring their own equipment. Each operator has great agency and influence on the effectiveness of the industrial workforce. Each has thousands of hours of field knowledge from sitting next to the problem, but turning that expertise into systems took an engineer. Traba’s operational systems improved only as fast as engineering bandwidth allowed. Coding agents changed that.

We didn’t have to send forward-deployed engineers to the problem: the people already sitting next to it now had the potential to build.

The legendary chef Auguste Gusteau of Pixar’s Ratatouille has a simple motto: ‘Anyone Can Cook.’ A soul brimming with creativity and motivation is limited never by their raw talent, but only by their discipline, desire, and determination. We wholeheartedly agree.

If we believe that Anyone Can Cook, what are the tools we need to provide for each person to become a Michelin star chef?

It took four things:

1. A kitchen. An agent on every laptop, with one-click onboarding and setup.

2. A cookbook. Recipes anyone can pick up.

3. A menu. A path for the best dishes to reach everyone.

4. A critic. Someone who tastes every idea before the cooking starts.

We’d like to say we designed all four in advance. We didn’t—the cooks showed up first, and we built the kitchen around them. Shortly after we handed out the agents, the operators built so much, so fast, that each of the other three became urgent ahead of schedule.

In the last 6 months, nearly half of our operators have used our internal builder platform, cooking a collective 100+ projects that turn their learnings from the field into automated systems. And all those operators building their own tools unlocked a ton of engineering horsepower to point at even harder problems…

Getting into the kitchen

Even from inception, this initiative was born out of necessity. Early in the year, our most experimental operators started hacking together quick apps with their agents and hosting them wherever they could: personal Railway and Vercel accounts, a free Cloudflare page, etc. One shipped a whole Slack bot alone, and we only found out when it outgrew the free plan on a personal account. Agent usage was climbing every week—either we let chaos reign, or we had to set some house rules, fast.

By March the whole org was on Claude Code, and we rolled out a default setup for everyone’s agents, installable in just 2 minutes via a single prompt. We called it Prometheus, and we told operators their mission remained the same: keep going out into the field, learn what works, and bring those operational learnings back into the core product. But now, they could prototype as fast as they could learn.

Inside, Prometheus carried what a veteran engineer would want a rookie builder to know before shipping anything: correct authentication and secrets management, toolchain recommendations and general development best practices, even the Traba design system. All deploys had to pass the pre-flight checklist, a mix of deterministic CI checks and agentic analysis. And knowing one day an engineer might pick up a prototype that had become an invaluable operations tool and rebuild it properly into the core, we had the agents document the operators’ brainstorming and product thinking and decisions, so projects could not only come to life quickly but also readily evolve into core production systems, or be pruned gracefully. All of this was fully agent managed.

Then we gave our pioneers a stage. During weekly All Hands they would share the fruits of their agents’ labors as those projects began to yield outsized results.

Exhibit A: Fahad began at Traba on our Live Ops desk and went on to lead Biz Support, the team that handles our business customers’ requests. Biz and Worker support was a 30-person team across five countries, forecasting upcoming support needs based on fluctuating regional markets and then assigning shifts around the clock to provide 24/7 coverage for our marketplace. Our support teams run on precise, painstaking coordination, and Fahad knew we had to solve this Sisyphean task. So, over the course of two weeks he built out a scheduling system that could forecast and draft a week of shift assignments in under 10 seconds.

The app kept growing from there. Karl, who leads our worker support team, grafted Cert Check onto it to auto-verify the certification photos workers send in: he filtered our historical data into offline evals and kept human-in-the-loop grading for difficult cases, saving an estimated 2 operator-weeks per month, for a ~95% reduction in operational expenditure for that task.

One operator’s project to streamline his team’s scheduling became the de-facto SaaS tool for Traba’s entire marketplace support pillars.

Operators began to snowball on each other’s work, and nobody had to wait on an engineering roadmap anymore.

Every lesson from those early builds went back into Prometheus, so month by month the process became frictionless. When Elijah’s onsites portal needed to run overnight with nobody logged in, Prometheus learned to give agents safe, scoped access of their own, clearing the path for every project that needed it next. As an increasing number of operator-built tools communicated with workers, we instructed Prometheus to route those requests through our central comms broker for proper batching and prioritization.

But Prometheus was a static config: a couple of engineers force multiplying operators, and soon operators had lessons of their own to share. That raises the next problem: once a cook makes something good, how does everyone else get to sample it?

Sharing recipes

Aside from applications, many of our operators naturally leaned toward building skills that they could easily invoke in Neutron, our internal operations agent, or in their own agents to trivialize a formerly complex area of work. In August we formally opened Bazaar, our internal marketplace for skills. Anyone can publish a skill, and anyone can subscribe to a set of skills. The catalog has since passed 200 skills.

Explore a sample of Bazaar:

People naturally subscribe to their teammates’ skills. When a skill becomes a team staple, the department head curates it into the team’s cookbook. Required governance and game-changing tooling go in the house cookbook.

The most popular by a country mile is the Slack bridge: a skill that initializes a basic personal bot running on top of existing agent harnesses like Claude Code and Codex. We encourage everyone to run and customize their own bots, and the earliest adopters have abandoned CLIs and IDEs and live in Slack, coordinating personal bot legions specially tuned to them: custom comms layers, routing, and remote environments.

Curating the menu

Recipe or app, the best dishes go the distance and make it onto the menu: proven in the field—battle-seasoned some might say—and folded into the core product. The goal is not for every prototype to survive natural selection: too high a hit rate means we aren’t ambitious enough; too low implies a lack of focus and execution.

In April, Fernando hacked together a small web app that turns a worker’s history into a resume formatted the way our customers ask for. It was live for the team the next day, and over the next few months, Fernando kept cooking, and adoption climbed: PDF parsing, AI bullet points, call-transcript-to-resume direct translation.

By June, it became clear that the resume tool deserved a spot in our core Operations Console. Fernando and a single engineer ported it over one afternoon, and it shipped for every operator. Our worker-application automation soon picked it up automatically.

Not every dish lands in the console. As operators built their own tools and tuned their agents to their own workflows, engineers laid the underlying agent platform and infrastructure to turn all that chaotic experimentation into its final form: Neutron. Neutron deeply understands our vertical, reasons over all the operational knowledge and signal we’ve collected the past 5 years, and takes actions in our systems. Every prototype that survived signaled to us some gap in Neutron’s capabilities, and the most successful experiments our forward-deployed operators run get baked into Neutron directly.

It’s not just projects that get promoted. Operators do too, as they become increasingly technical. Will started as a technical operator on our Scaled Systems team, automating service for the long tail of light industrial businesses we work with. Armed with coding agents, he shipped a worker interview scheduler that earned an All Hands spotlight in April. He sent his first pull request to the core codebase that same month and has since transitioned into a full-time engineer with hundreds more merges under his belt.

A critic in every kitchen

By July, the kitchen hit critical mass. The number of people committing code to non-engineering projects nearly doubled from June to July, and commits more than doubled.

With dozens of people building at once, nobody knew what else was being prepared in the kitchen; in late August we counted seven different candidate-pipeline trackers, two of them for the same customer site and built just nine days apart.

Anton Ego is the most feared man in Ratatouille’s Paris. He samples every dish and holds each one to the highest standard for the gourmands of the city. We needed our own Anton Ego—something to (constructively) critique every chef and every creation.

A good product manager asks a handful of questions before anything gets built. Why are you building this? Who is it for? Where should it live? Has someone already built it? One PM can ask those questions a few times a day. Prometheus could tell cooks not to take an idea at face value, and to understand the problem space and iterate and pressure-test like any product-minded engineer would, but it couldn’t make connections across the entire org and flag conflicts. Enter Oracle, our AI product manager that sits inside every cook’s sous agent and critiques the idea before implementing.

Oracle pauses the agent and scours the company’s collective knowledge: our code, our other deployed apps, our tickets and roadmaps and internal docs, and what Neutron can already do. Then it gives its verdict: either this already exists so use that tool, or something close exists so extend, or this is novel and worth pursuing.

Our critic differs from Ego in one important way: it can’t close the restaurant. “Build it anyway” always works. When two operators attack the same problem differently, we want both of them cooking so we can taste which recipe turns out better. Oracle’s job is to give each cook constructive feedback and the context of what the rest of the kitchen has already tried, so a parallel experiment produces a real signal instead of a duplicate.

Every check also tells us what operators actually need. Oracle scores each ask for business value and records what blocked it, which turns the critic into a running list of the most acute pain points on the ground floor. In the last month, across hundreds of conversations and asks, it’s flagged and clustered several as the most pressing needs: missing data, connections to another system, new tools for Neutron. That list comes unfiltered from the field and informs our roadmap.

The critic’s palate is about to get broader. It sees the explicit asks but not the hundreds of small tasks operators hand their agents every week to complete in a variety of ways. This month, Oracle starts learning directly from those agent sessions to accelerate Neutron’s development even more. Today, Neutron improves as operators teach it, but tomorrow it’ll be automatically learning from how our best operators work.

From anywhere

Hayek called the economic problem of society “mainly one of rapid adaptation to changes in the particular circumstances of time and place,” and concluded that “the ultimate decisions must be left to the people who are familiar with these circumstances.”

At the end of Ratatouille, Ego tastes a plate cooked by a rat and writes the review of his life: “In the past, I have made no secret of my disdain for Chef Gusteau’s famous motto, ‘Anyone can cook.’ But I realize, only now do I truly understand what he meant. Not everyone can become a great artist, but a great artist can come from anywhere.”

We still wholeheartedly agree, and we think it’s how a startup survives: the fittest are the ones that adapt fastest. At the start of the year, we asked ourselves how we could give massive agency to each operator on the ground floor. Today we’re launching and ramping at enterprises we weren’t sure how we could service back then, and we’re accelerating through a fundamental shift in how Traba improves. Where operators once accumulated knowledge that engineers could only partly encode, they now rapidly improve our systems directly, with engineering’s support and guidance. Somewhere in the field, the next great dish is already cooking.

We’re hiring

If you want to build the kitchen a whole company cooks in, come join us.