Factory and Cognition Ship Rival Tools to Run Agent Fleets

Factory and Cognition, two companies selling AI coding agents to engineering teams, each shipped a way to run many agents at once within about eleven hours. Factory made Environments generally available on 12 October 2026, and Cognition’s Devin account announced Dynamic Workflows on 13 October.

Factory says Environments are reusable sandboxes for its Droids, rebuilt every few hours. Cognition’s documentation says a Dynamic Workflow is “a deterministic Python script that orchestrates a team of Devin agents”.

What Factory Shipped

In its post “Droids now have a fleet of sandboxes”, Factory says: “An Environment is a reusable definition of the sandbox your Droids run on, composed of the repositories, tools, and dependencies they need.” It builds them into portable images ahead of time, “so each sandbox is ready before the Droid arrives”.

The post says “Factory rebuilds Environments every few hours and validates each new build. Only successful builds are promoted, so Droids never start from a faulty environment.”

Factory ties the launch to its long-running work. It says Missions, “our long-running multi-agent framework, work best on repositories at level four or above” of its five-level Agent Readiness score, and that “most enterprise customers start at level two”. With Environments, it says, “you can fan out the fixes across multiple sandboxes.”

Environments “are available today for all plans”, the post says. It calls this the start of a week of launches: Droids on “the hardware your team already owns”, on “sandboxes that never leave your infrastructure”, and “in the places engineering work begins”.

What Was Posted on X

Factory’s own launch post of 12 October 2026 (19:09 UTC). It says the company is shipping its vision for the best cloud agent experience this week, starting with the sandbox where an agent runs, and introduces Environments as generally available.

The Devin account’s post of 13 October 2026 (06:02 UTC), about eleven hours later. It says Dynamic Workflows let Devin split a big job into pieces, hand each one to its own Devin and bring the results back together, with the plan living in code.

Factory’s post of 19 October 2026 (16:25 UTC). It says Factory is partnering with Databricks so its Droids can build, migrate and secure software directly on Databricks; a follow-up post says customers with universal Databricks commitments can apply them to buy Factory on Databricks Marketplace.

Cognition’s post of 19 October 2026 (16:46 UTC), 21 minutes later. It calls blanket usage caps “a hiring freeze in disguise” and introduces efficiency-based auto-approvals for a customer’s most productive engineers, plus Devin Coach to help everyone else get there.

What Cognition Shipped

“Dynamic workflows let Devin split a big job into pieces, hand each one to its own Devin, and bring the results back together,” the @devindevelopers account posted. “The plan for a Dynamic Workflow lives in code instead of an agent’s head.”

Cognition’s documentation says Devin writes the script, runs it, and the script decides which agents run, in what order, and what each one is told. By default each agent “is an independent Devin session on its own VM”.

The docs recommend a workflow for “roughly five or more independent units” or a staged pipeline such as “audit → fix → verify”. Runs are recorded and can resume where they stopped, and “The default and maximum budget for a run is seven days.”

On cost, the docs say: “Every agent in a workflow is a Devin session, so one run can consume far more ACUs than doing the same task in a single session.” Enterprise accounts must have an admin turn the feature on.

Two Routes to the Same Job

Both launches are about long-running work spread across many agents. As we read the two texts, Factory starts from the machine, a validated sandbox that every Droid can boot into, and Cognition starts from the plan, a script that decides which Devins run and in what order.

Factory’s chief technology officer, Eno Reyes, posted that the company “looked across the market at every cloud agent product to identify the missing gaps” and listed seven design goals, including “API-first so environments can be spun up by agents quickly and easily” and enterprise controls “to enable 10,000+ users to benefit”.

Sequoia partner Shaun Maguire, replying to Factory’s launch post, wrote: “Gonna be a good two weeks!”

The Rivalry Behind the Launches

The two companies have been in a public dispute since 30 September. Factory CEO Matan Grinberg posted that day: “We are terminating Chris Degnan for unethical conduct involving Cognition.” He called Cognition “A much larger competitor”.

Degnan replied the same day: “You did not terminate me. I resigned from my advisor position on Monday and told you I was going to Cognition.” He added: “Cognition has never asked about information on Factory, nor would I ever give it.”

On 25 September Cognition posted that it “crossed $1B in annualized revenue run rate”. This publication covered Factory’s $200 million raise at a $5 billion valuation and how the two companies bet on opposite sides of the AI stack.

Update, 19 October: Factory Sells Through Databricks as Cognition Ties Devin Budgets to Efficiency

Both companies announced new enterprise features on 19 October, 21 minutes apart on X. Factory said it is partnering with Databricks “so customers can use the data and models in the Databricks Data + AI Platform to power software that is built or migrated by Factory”. Its post lists three uses: PDLC Automation, Migration and Modernization, and Cybersecurity, the last to “remediate vulnerabilities in the applications monitored by Databricks Lakewatch”. It also says “qualifying customers can purchase Factory through the Databricks Marketplace using up to 10% of their existing Databricks commitment”. Factory says the two companies are helping “one of the largest global enterprises”, which it does not name, migrate a data platform with “petabytes of data and hundreds of thousands of tables”.

Cognition introduced three Devin features for enterprises. An Efficiency Score, a 0–100 rating, grades each engineer on “productive hours per ACU”, Cognition’s unit of spend, with a smaller credit for volume. Devin Coach gives engineers feedback on a prompt before it runs. With efficiency-based approvals, “When an engineer with a Healthy or Satisfactory grade nears their limit, Devin raises it automatically, up to a ceiling the admin sets”, while other engineers’ requests go to an admin. Cognition says that at one company with auto-approvals the cap was lifted 22 times and the engineer went “from 30 to nearly 80 merged PRs a week”. Its thread names Bilt as a user and points to a story in The Information asking why a company that charges by usage would help customers rein it in.

As we read it, both moves go after the same enterprise budget line from opposite ends: Factory lets buyers pay with money already committed to Databricks, and Cognition ties how much an engineer may spend on Devin to measured output. Neither company’s announcement mentions the other.

The Structural Read

Factory’s post frames the sandbox as onboarding. An agent in a new environment, it says, has to clone repositories, install toolchains and find undocumented configuration before doing real work, spending “time and tokens onboarding to your codebase rather than completing meaningful work.”

Cognition’s docs frame the problem as coordination. In a workflow “the orchestration itself is code”, and every agent call is recorded, so a run can be watched while it executes and resumed if it stops.

As we read it, each company is building the layer around the model that its customers pay for: the environment in Factory’s case, the orchestration in Cognition’s. Both texts point the work at long jobs that run while engineers are away.

Three Implications

ENGINEERING TEAMS Factory says most enterprise customers start at level two of five on its Agent Readiness score, and that Missions work best at level four or above.

BUDGET OWNERS Cognition’s docs say one workflow run can consume far more ACUs than a single session, and recommend testing on a slice first.

THE TWO COMPANIES As we read it, the launches put two direct competitors in the same feature race within a day, while their public dispute of 30 September remains unresolved in the posts we read.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework puts it this way: “The surface, not the model, governs autonomy.”

As we read it, both launches are bets on that surface: Factory on the sandbox and its policies, Cognition on the code that coordinates many Devins. Neither text is about a new model.

What Is Not Established

Neither launch text we read gives a price for the new feature, a customer name, or usage numbers. Factory says Environments are on all plans; Cognition says workflow runs consume ACUs per agent session, without a price per run.

Neither company’s launch text names the other. The account of the 30 September dispute rests on the two men’s own posts; we have not seen a court filing.

The Bottom Line

Within about eleven hours, Factory made Environments generally available, reusable sandboxes for running many Droids in parallel, and Cognition launched Dynamic Workflows, code-defined plans that split a job across many Devins. Both pitch long-running work spread across agent fleets; neither text gives a price, customer or usage figure.