The short answer is that Jobbit is a UK company with one loop viewed from three angles. Jobbit.uk is where work gets assigned and finished, by an AI agent and by vetted people. Jobbit Labs is what that work teaches. Jobbit Security keeps the whole thing, and our customers' systems, tested against the people who would like to break it. The company published its own long-form answer this week, What Is Jobbit? Inside the UK AI Agent and Human Network, and this post is the Labs view of the same story: what the platform is, and why it is the data engine underneath everything we do here.

The company in one table

One company, one team in London, one registered office at 124 City Road. The websites are separate because the customers are, not because the work is.

Jobbit.uk: the agent and the human network

Most people meet Jobbit through a chat box on jobbit.uk. You describe a job in plain language and the platform routes it to the right agent, workflow, file generator, app builder or expert flow. The output is real work: a hosted web app, a scheduled automation, a document or spreadsheet, a research report with citations, an image or a video, or one of the free online tools that turn an everyday chore into a single message.

The part that matters most to Labs is what happens when software is not enough. The agent can hand a job to a person. Jobbit Pro is the human side of the same platform: vetted freelancers and specialists, from developers and designers to tradespeople and consultants, who take the jobs an agent should not do alone. The agent drafts the brief, the expert delivers, payment sits in escrow until the work is accepted, and the record of the whole job stays in the same workspace as everything else. The company's guide to the 12 jobs an AI agent can do for a business this week is a good picture of the agent side; the human side is what makes the physical world reachable at all.

Why a research lab sits inside an agent company

Frontier models were trained on the internet. The internet has no record of how a boiler gets serviced, how a mixed order gets picked from a live warehouse, or how a fibre line gets spliced. Nobody wrote those tasks down step by step, with the false starts, the corrections and the verified result. That is the data gap Jobbit Labs exists to close, and it is exactly the gap the agent platform runs into every day when a task needs a person.

Every task on Jobbit runs through the same four stages, and the stages are the company in miniature.

1. Assign. The brief is written or refined by the agent and routed to the right tool, agent or vetted human.
2. Execute. The work is done: code written and deployed, the document produced, the job completed on site.
3. Record. The instructions, the plan, the actions, the corrections and the verified outcome are captured inside the platform, with the customer's consent.
4. Improve. Those records become the training and evaluation data for the next generation of agents, and for the world models Labs is building.

Anyone can bundle tools. Almost nobody owns the loop. Because the agents and the network they book both belong to Jobbit, every task runs inside one system and stays there. That closed loop is what Labs is built on, and it is why the data has properties you cannot buy:

• Real trajectories, not summaries. Brief, plan, actions, mid-task uncertainty and the verified outcome, recorded as the work happens rather than tidied up afterwards.
• Multimodal by default. Text, photos and video of physical work sit alongside the chat, files and app builds of knowledge work.
• Human and agent, side by side. The same tasks performed by AI and by vetted experts, with comparable outcome labels, which is what you need to measure where the machine is good enough and where it is not.
• Provenance and consent built in. Captured inside our own platform with clean lineage. Nothing is scraped, and customers opt in.

What Labs does with it

Three things, in increasing order of ambition.

Datasets. First-party records of how physical and knowledge work is actually done, packaged for AI and robotics teams that have plenty of compute and almost no data of this kind. The use-case explorer on our home page shows the six task families we started with, from warehouse picking to weld inspection, and the signals each one captures.

Evaluations. Because the same task is performed by agents and by people, with the outcome verified either way, Labs can score an agent against a human baseline on real work rather than on a benchmark written for the purpose. The data loop section explains how a completed task becomes a label.

World models. World models predict the outcome of an action in the physical world before it is taken. They do not predict the next video frame pixel by pixel; they learn in an abstract space, encoding the current state, predicting where an action leads and correcting themselves against what actually happened, the principle behind joint-embedding predictive architectures (JEPA). Recorded trajectories with verified outcomes are the raw material, and our world models page has an interactive demo of why more of that data means smaller prediction errors.

Labs is early, deliberately. The orchestration layer that routes every Jobbit task and labels each completed one runs live on jobbit.uk today, and so does the capture loop. The research is built on top of production, which is the difference between a dataset you can buy and one you can trust. We do not publish numbers we cannot stand behind, so you will not find success rates on this site until they mean something.

Jobbit Security: the red team down the corridor

The third division is easy to explain from where Labs sits. The agent platform and the data engine handle customers' code, files and operations, and we would rather be tested by our own red team than by someone else's. Jobbit Security runs manual penetration testing, cloud security assessments, red team operations, phishing simulation and vCISO work for UK businesses, aligned to OWASP, NIST SP 800-115, PTES and MITRE ATT&CK, and it tests Jobbit's own systems first. The same AI that lets a business build an app in an afternoon lets an attacker find that app's weaknesses in an hour, so a company that builds with AI should be able to defend with it.

What this means for you

• AI and robotics teams: talk to us about datasets, evaluations and world models built on first-party, consented real-world data. Start with the use cases and get in touch.
• Enterprises with physical operations: the tasks your teams do every day are the data frontier models lack. Recording them, with consent, is a partnership rather than a purchase.
• Researchers and engineers: we are hiring in London. The careers page lists the roles.
• Businesses and freelancers: the fastest way to understand the loop is to run one task through it. Give Jobbit a job on the free plan, no card needed, or join the human network at pro.jobbit.uk.

Read the full story

The complete company page, with pricing, principles and who is behind Jobbit, lives on the main site: What Is Jobbit? Inside the UK AI Agent and Human Network: Jobbit.uk, Jobbit Labs and Jobbit Security. Its own boilerplate says it in three sentences:

Jobbit is a UK AI company bringing AI to the physical world. Its platform at jobbit.uk combines a multipurpose AI agent, which builds apps, documents, research and media, with a network of vetted human experts who finish the work an agent cannot. Jobbit Labs builds first-party datasets and world models for physical AI, and Jobbit Security provides penetration testing and red teaming for UK business.

If you only remember one line from this page, make it the company's own: finished beats generated. Everything Labs records starts with work that was actually done.

For the people who use them, the point is obvious: a job that used to mean finding a website, dodging its upgrade prompts and hoping the watermark is small now takes one click. For Jobbit Labs the interesting part is different. These are the smallest complete tasks on the platform, and the cleanest record we have of the moment a person hands work to a machine.

What a one-message task looks like

Take the QR code page. The visitor pastes a link, chooses a size and a format, and proceeds. Behind the button, the page writes a message such as "Generate a static QR code for this link, 1000 by 1000 pixels, PNG and SVG" and gives it to the agent. The agent generates the code, checks it and returns two files. The whole exchange is one message in and one result out, which is why the company calls this the pattern it wants every new user to learn first. The guide to free QR codes that never expire shows the exact message and explains why static codes are the ones to print.

The other three tools follow the same shape:

• Compress a PDF. Upload a file, pick a target such as under 1 MB, and the agent shrinks it while keeping the text sharp, with no upload limit. The PDF guide covers the size-versus-legibility trade-off.
• Resize or convert an image. HEIC to JPG, under 200 KB, or one of 14 platform sizes for social profiles and listings. The image guide lists the dimensions.
• Write an invoice. Client, amount, terms, and the agent produces a UK invoice with the nine fields a finance team expects, including late-payment terms. The invoice guide explains each field.

The company's overview, QR codes, PDF compressor, image resizer and invoice generator: four free online tools, links all four pages and guides in one place.

Why tiny tasks matter to a research lab

Labs spends most of its time on the hard end of the spectrum: how a warehouse pick is done under time pressure, how an engineer diagnoses a fault, how a weld is judged. Those tasks are rich, but their outcomes are expensive to verify. A four-second tool task is the opposite, and that is precisely its value.

The outcome is objective. A QR code scans or it does not. A PDF is under the target size with legible text or it is not. An image has the requested dimensions and file size or it does not. An invoice either contains the nine required fields or it is missing one. Checking these results costs almost nothing, and there is no room for a generous grader.

The request is short and the intent is clear. A one-line message with a handful of parameters leaves little to interpret, so when the agent gets it wrong the failure is informative rather than ambiguous.

The volume is high and the variance is low. Thousands of near-identical tasks give a stable baseline. When the agent's underlying model changes, this is the first place a regression shows up.

They are the same loop as everything else. Assign, execute, record, improve. A one-message tool task exercises the whole pipeline that a forty-page tender response or an on-site job does, in miniature, which makes it the natural smoke test for the platform Labs builds on. Our post on what Jobbit is and where Labs fits explains that loop in full.

What each tool teaches the agent

Each of the four exposes a different skill, and each skill generalises well beyond the tool.

The invoice case is the one people underestimate. UK late-payment rules give a business the right to statutory interest at 8 per cent above the Bank of England base rate plus a fixed sum of £40, £70 or £100 depending on the debt, and an invoice that states the terms correctly gets paid sooner. An agent that produces that clause reliably has learned something a template cannot teach.

From a form to a trajectory

What Labs actually works with is not the file. It is the record of how the file came to be, captured inside the platform with the customer's consent: the message, the agent's plan, the actions it took, the result, and above all the corrections. "Make it bigger." "That is still over 1 MB." "Add my VAT number." Corrections are where the signal lives, because they show the gap between what the person meant and what the agent understood, in the person's own words.

Those records do three jobs. They give the agent team a regression suite drawn from real use rather than invented prompts. They give Labs a measured, objective baseline for the wider question of how reliably agents turn language into verified results. And they show how people learn to delegate: what the second and third messages in a thread look like once someone has seen a one-message task succeed.

What we will not claim

We are not going to quote success rates for these tools on this site. The numbers exist, they are watched, and they will be published when they mean something outside our own dashboard. Labs is early by design, and the fastest way to lose the trust of the AI and robotics teams we work with is to dress up a launch-week figure as a finding.

Try one

The tools are free, there is no card, and there is no watermark. Start with the one you needed last week:

• Make a QR code that never expires
• Compress a PDF to under 1 MB
• Resize or convert an image, including HEIC to JPG
• Write an invoice with the right UK terms

If you run a team that would rather measure agents than take our word for it, that is what Labs is for. Get in touch.