Skip to content

One operator, many lanes

How Runway runs a business's real work with AI agents: highly capable, with guard rails, and your yes on everything that matters.

Most AI agents still wait for someone to type. Ours don't. Runway reads a business's email, chat, tickets and meetings, turns what it finds into work, and does it: replies written, pages built, code shipped, invoices issued.

It does each of those as soon as the right person says yes. That's the design: a team of agents that can do almost anything, inside clear lines for what needs a person.

Done well, it feels like 20 employees in the palm of your hand, with a lot less on your mind.

The layer above the agents

An agent harness is everything around the model: the loop that lets it use tools, its memory and what it's allowed to touch. Some companies build that loop from scratch. That's a fair bet if you sell agents inside your own product.

We made a different bet. The agents you can buy get better every month, so we don't rebuild them. We run the best models from three providers and switch when a better one ships. What we built sits on top: who does what, what each agent may touch, what it has to prove and when it has to stop.

How it works

Each workspace has one coordinator. It watches the places work arrives, decides what matters and hands each job to a short-lived agent with one task. We call those agents lanes, and many run at once.

  1. Watch. Email, chat, tickets and meetings become one list of work.
  2. Work. Each job goes to a lane, on the model that fits it.
  3. Prove. A review by another provider's model, then a live check.
  4. Ask. Anything that needs a person becomes one short card.

Every lane ends with a receipt: what changed, what shipped, the live check and anything still waiting on a person. Nobody has time to read agent transcripts. Anyone can read a receipt in a minute.

A stuck lane says so and names the blocker instead of guessing. Fixes move up to stronger models. If a problem keeps coming back, the agents stop and hand a person a short list of what they tried.

The rules are plain text that every agent reads before it acts. Change a rule today, and the next agent that starts follows it. Runway runs on cloud systems that stay on around the clock, and it grows by adding machines.

Highly capable, with guard rails

Runway sends the email, posts the update, merges the change and issues the invoice. It does each one as soon as you approve it, with the exact words, recipients and amount you saw.

  • On its own. Research, drafts, code, tests, reviews, pages and reports: work you see before anyone else does.
  • On your yes. Email and messages to people, merges and releases, anything that spends money or can't be undone.

Most of our design went into the asking. A question arrives as a card you can answer in seconds, with a recommended answer, the facts behind it and a link to the exact thing. The agents keep working on everything that doesn't depend on it.

68% of people don't have enough uninterrupted focus time in the workday, and 64% struggle to find the time and energy to do their job (Microsoft Work Trend Index 2023).
Source: Microsoft Work Trend Index 2023, "Will AI Fix Work?"

Agents never hold your passwords or keys. The tools that need one use it without showing it. Every action keeps a record of who approved it and when.

What it's worth

The value is time back: hours you no longer spend sorting, drafting, chasing and checking. Your numbers depend on your work, so here's the math instead of a promise.

Example math: hours back each week times hourly cost times 48 working weeks. 2 hours a week at $50 an hour is $4,800 a year; 5 hours at $100 is $24,000; 10 hours at $150 is $72,000.

Find the row that sounds like your week. The part no table shows is the quiet: fewer open loops in your head at the end of the day.

Tenancy: your workspace is yours

We run work for many businesses at once, so tenancy is built in from the start. Each tenant, whether a company, a team or one person, gets a workspace of its own.

  • Its own records. Queue, answers, sends and receipts.
  • Its own connections. Each tool connected by its owner, never borrowed.
  • Its own brain. What it has learned about how you work.
  • Its own approvals. Only the right person's yes counts.

Isolation is the default. One tenant's facts never show up in another's work. A workspace can't use another's sign-ins or send as anyone else, and every question has exactly one owner.

A brain that grows with you

Every answer you give is evidence. When the same choice keeps coming up, your workspace suggests it as a preference, like "send routine replies from my inbox," and you confirm it before it becomes how things work. A correction always beats an old habit.

The brain shapes how work gets done. It never adds authority: a learned preference can't send, spend or delete on its own.

For engineering teams: a runway for your repos

Writing code got fast. Reviewing it didn't, so that's where an engineering runway spends its effort.

Developers with an AI pair programmer finished a coding task 55.8% faster. Reviewers find the most defects reading 200 to 400 lines at a time; past 400 lines, the ability to find defects drops.
Sources: Peng et al. 2023, "The Impact of AI on Developer Productivity"; SmartBear, "Best practices for peer code review"

Every pull request becomes a question card built for a fast, careful review:

  • What changed, in plain words, with screenshots of the change on phone and desktop.
  • An impact score from 1 to 5 (coming soon): how much it can hurt if it's wrong. A copy change will score low; sign-in, data and money will score high.
  • Look at these first: the two or three spots that matter, with the code on the card, so you read 20 lines, not 2,000.
  • A recommended answer: Approve, Comment or Request changes, with the review comment drafted in your voice.
Example review card: impact 3 of 5, what changed, two code spots to look at first, phone and desktop screenshots, and Approve recommended.
Example: a review card built from one of our own changes, mocked up for this article. The impact score is coming soon.

Soon each team will set a rigor level from 1 to 5. At 1, passing tests will be enough. At 5, every change will also get an independent review by another provider's model, a live check and screenshots before it reaches you.

Your standup becomes a list of receipts, each with the commit, the test run and the live check. You keep the coding agents and subscriptions you already pay for, and you don't need a platform team to start.

We build Runway with Runway

Most of Runway was written by its own agents, under the same approvals and reviews our customers get.

The models will keep changing. The job stays the same: line up the decisions, get a yes, let the agents build and prove it, and keep the record. If you want to see that from an owner's side, line up a morning with us.