What Cognition launched

Cognition makes the Devin coding agent, and it launched Devin Bots on 17 November (AlphaSignal; Fortune). A Bot is a long-running agent that sits above ordinary Devin sessions. You give the Bot a goal and it splits the work into tasks. It then runs several Devin sessions in parallel and keeps each session going until its pull request is ready to merge, according to the Devin Bots product page. Devin's release notes list the feature under 19 November. They describe Bots as "always-on Devins with their own identity, memory and permissions."

A standard Devin session handles one task that a person hands over. A Bot owns a standing outcome instead, such as keeping CI healthy or clearing a ticket queue. The Bot contacts a developer when a task needs judgement or access it doesn't have (AlphaSignal). CI, short for continuous integration, is the automated build and test run that checks every change before it merges.

How a Bot runs the work

Each Devin session gets its own isolated virtual machine with a code editor, a browser and a shell (AlphaSignal). The Bot coordinates those sessions. It takes in work from configured sources such as Slack, Linear tickets, alerts, schedules and CI results, and breaks each objective into smaller tasks. Each task gets its own concurrent session. People can reach a Bot through the web app or Slack, by text message or phone call, or in a group chat (release notes).

Fortune reports that a Bot joins a team's Slack as a member and keeps context across weeks, and that several colleagues can work with the same Bot. According to Fortune, Cognition sends planning, design and computer-use steps to stronger models and routine code to cheaper ones, including models Cognition trains itself. Each Bot can also have its own budget, measured against its results. Fortune adds that the Bot technology builds on The Interaction Company, maker of the Poke assistant, which Cognition bought in July. Bots can pass work to one another as well. In the product page example, a QA Bot finds a bug and brings in a Developer Bot to fix it.

Templates and triggers

Cognition launched six templates. Each one packages triggers, instructions, schedules and handoffs (AlphaSignal). The product page describes them as follows:

  • Software Factory: looks for bugs, polish, performance and refactoring work in a named product area, opens one small PR per item and learns from what you merge or close.
  • Android Engineer: builds Android features in Kotlin and ships them to internal testing.
  • QA Engineer: tests each PR in a real browser, posts a video and reports pass or fail on the PR.
  • Latency Optimizer: finds slow endpoints, jobs and queries, then each week attempts a fix for the worst one and opens a PR with before and after timings.
  • Product Engineer: takes a ticket through to a tested PR and through review.
  • Chief of Staff: watches channels, posts updates and follows handed-off work until it's finished.

Schedules and triggers keep a Bot working between human messages. The product page's examples include watching Slack for customer issues, checking CI health each night and monitoring Linear for blockers.

Cost and access

Devin's pricing page lists "Access to Devin Bots" on two plans: Max at US$200 a month and Teams at US$80 a month. The US$20 Pro plan doesn't mention Bots, and Enterprise pricing is negotiated. Usage is metered as allowances that refresh daily and weekly, with costs that vary by model, task size and complexity. A Bot that runs many sessions in parallel will therefore use up an allowance faster than one interactive session. Cognition hasn't published a price specific to Bots (AlphaSignal).

Enterprise admins get firmer spending controls. Devin's usage policies documentation describes a budget tree up to three levels deep, plus monthly per-member limits measured in ACUs (agent compute units, Devin's billing unit). Once a budget or a member's limit is reached, new work stops on both local and cloud surfaces. Work resumes when an admin approves more ACUs or the monthly window resets. Bots start sessions on a schedule, so these caps are the main protection against a runaway bill.

On security, the product page says every session runs in its own VM. The page also lists RBAC, SSO, SCIM, private networking, audit logs and zero data retention with model providers. Fortune reports that admins choose which repositories and tools each Bot can reach. Network controls aim to block data leaks set off by prompt injection, which means malicious instructions hidden in content an agent reads. Bot code goes through the same tests as human code, and each team decides which changes need human review.

What's still unproven

All of the headline numbers come from Cognition or its customers. Cognition says beta testers merged about 4.5 times more pull requests per message than with direct Devin sessions. It also says some engineers went from about 50 to nearly 2,000 PRs a week (AlphaSignal). The product page quotes Bilt's VP of Engineering saying Bots merged about 1,600 PRs in a single week. AlphaSignal points out that Cognition hasn't published sample sizes, median results, defect or rollback rates, review time or compute costs. Fortune reports that Bot usage grew tenfold in two weeks across dozens of early-access customers. That figure shows adoption and says nothing about code quality.

PR volume is a weak measure for a tool built to produce PRs. A Software Factory Bot that opens one small PR per finding will post large numbers while moving the work onto human reviewers. Cognition also hasn't said how a Bot settles conflicts when parallel sessions edit the same code.

The competition matters for anyone choosing a platform. Crypto Briefing dates OpenAI's Dots to about 29 September and Anthropic's Duet to about 21 October. Dots plugs into Codex and Duet connects to Claude Code. Cognition argues that Bots are built around a coding agent from the ground up. Fortune notes that Cognition relies on the same model providers it now competes with.

What New Zealand builders could try

All plan prices are in US dollars, and Bots run on Cognition's cloud machines. Teams whose client contracts restrict where source code goes should check those terms before connecting a repository. For a small team, the QA Engineer template is a contained first trial. It works on PRs that already exist and reports pass or fail without merging anything. A team could run that template for a few weeks on one repository with good test coverage. Measuring reviewer time, revert rate and allowance burn would give a better answer than any vendor PR count. Teams on Enterprise should set budget and per-member caps before turning on any scheduled triggers.

The Software Factory and Latency Optimizer templates suit codebases with a backlog of small, well-tested fixes. Those templates fit poorly where changes need domain judgement that a ticket can't capture. If you run a trial and measure the results, you can share them on our projects page.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member