Skip to content
AI Primer
release

Cognition launches always-on Devin Bots for ongoing goals

Cognition's Devin Bots pursue ongoing goals by launching coding sessions. Templates cover backlog, CI and releases; Cognition says a bot swarm maintains its Android app without human Android engineers.

6 min read
Cognition launches always-on Devin Bots for ongoing goals
Cognition launches always-on Devin Bots for ongoing goals

TL;DR

  • Devin Bots manage ongoing engineering work by launching teams of coding agents, handling planning, review and testing, according to the launch thread.
  • Each Bot keeps its own identity, memory and goals across sessions, as swyx's technical follow-up describes.
  • Cognition says a Bot swarm builds and maintains its Android app without human Android engineers, in its Android app demo.
  • Beta testers merged 4.5 times more PRs per message on average, according to Cognition's announcement.

Devin's own account introduced its new “army” of Bots; the Android Bot merged 1,134 PRs from 183 human messages. A studio manager Bot collects decisions four times daily and accepts emoji replies. Everyone with access to a shared Bot can also read its memory files.

Persistent memory and child sessions

Cognition launched Devin Bots on November 17. Each Bot investigates an assigned responsibility, delegates work and evaluates the results to decide what happens next.

The Bot documentation separates the manager from its workers:

  • Bot: A persistent chat, instructions and memory that carry across workstreams.
  • Child sessions: Individual pieces of work, each running on its own machine. The Bot starts and monitors them.
  • Separate modes: The Bot's own mode and its default child-session mode can differ. An individual child can also receive a different mode.

Persistent identity is the clever part.

Ongoing goals and automations

An assignment can be as specific as “maintain p99 latency below 100 ms for every endpoint,” using the example in russelljkaplan's explanation. The Bot would:

  • Launch sessions to investigate requests outside the latency SLA.
  • Periodically start proactive efficiency work.
  • Continue pursuing the target as the application changes.

Bots can also create, maintain and update the automations behind an outcome, as a follow-up explains. The product page gives concrete triggers: customer issues in Slack, nightly CI checks and blockers in Linear.

Engineering templates

Templates cover backlog work, CI and releases, with a blank-Bot option for custom responsibilities. The public catalog describes six engineering roles:

  • Software Factory: Finds bugs, polish, performance and refactoring work; opens small PRs with proof and learns from what gets merged or closed.
  • Android Engineer: Takes Kotlin features through tests and an installable internal build.
  • Chief of Staff: Watches channels, delegates work and follows it through completion.
  • Latency Optimizer: Tries a weekly fix for a slow endpoint, job or query, with before-and-after performance measurements.
  • Product Engineer: Takes requests through implementation, testing and PR review.
  • QA Engineer: Tests PRs in a real browser and posts video evidence with pass/fail results.

Needs You and bot-to-bot handoffs

Approvals accumulate in Needs You, with four documented item types:

  • Decisions, with a recommended option.
  • PR review or merge requests.
  • Permission requests, including access, automations, network hosts and usage-limit increases.
  • Reports and FYIs.

Bots can consult one another before escalating questions to a human:

Both Bots already had access to the relevant material; imjaredz said the other Bot might have the answer faster because it was working on something similar.

The available communication surfaces include Slack, text, phone calls and group chats, alongside the Devin app.

Android app in 44 days

Cognition says it had no Android engineers when Samyok, an engineer on its Bots team, assigned Android Bot the job of creating the app. Its 44-day account reports:

  • 1,134 merged pull requests.
  • 344,651 lines of Kotlin.
  • 2,841 screenshot tests.
  • 1,217 closed tickets.

The Bot filed tickets, launched a session for each and merged PRs that passed CI and Devin Review. Daily Slack reports included screen recordings.

Human involvement included setting direction and constraints, resolving blockers and handling decisions such as security reviews. Samyok sent no messages on 22 of the 44 days.

The app is in open testing on Google Play. Its ongoing assignment is feature parity with iOS.

Audio regression testing

LocalFlow's release checks had never actually recorded audio, allowing version 0.0.23 to ship broken, according to imjaredz's account. During follow-up maintenance, the Bot added a check that failed on the old code and passed on the fix, then ran the application on a Mac.

Layout Bot's three-hour checks

Cognition also disclosed a Layout Bot responsible for keeping its web app from shifting or lagging. Every three hours, it builds the latest production code and runs the full stack on its own machine.

Its checks cover:

  • First visits and repeat visits.
  • Page loads with individual requests deliberately delayed.
  • Reruns to distinguish regressions from timing noise.
  • Tracing failures to the introducing PR, followed by a fix with before-and-after recordings.

Cognition reports merging 159 Layout Bot PRs over three weeks and cutting lag by 70%.

Custom Bots and Unity testing

Nader Dabit, who works at Cognition, described his own Bot setup as six responsibility areas:

  • CTO: Cross-repository tasks, session access and status reports from other Bots, reachable through iMessage.
  • Auto(mation)Bot: Internal dashboards, recurring reports and automation performance.
  • DevCommunityBot: Ambassador-program data across spreadsheets and Notion, plus visualization mini-apps.
  • Poke: Personal assistance and non-engineering research.
  • DevRelBot: Documentation, tutorials and demo videos recorded from the actual product UI, with ElevenLabs-generated narration.
  • Gaming Bots: A team coordinated through SuperJenny for Windfable, a Unity RPG.

For Windfable, each Bot gets macOS, Xcode and Unity. Dabit says changes return with gameplay videos and side-by-side screenshots, including checks for folded-phone and unfolded-tablet layouts on the iPhone Duo.

A Windfable session with its macOS computer and iPhone Duo simulator

Access, permissions and usage

Bots are available now, and their work counts toward existing plan usage like other sessions, according to the documentation.

Access has several distinct boundaries:

  • Private by default: A new Bot remains private until shared.
  • Bot-owned credentials: Connections and secrets belong to the Bot, separately from the people messaging it.
  • Explicit grants: A Bot requests approval before receiving a person's sign-in, secret or session; the approval card describes what will be shared.
  • Administrative controls: Enterprise roles separately govern Bot use, organization-wide management and enterprise-wide management. Management permission alone does not grant conversation access.

Cognition's enterprise security claims include zero retention with model providers, audit logs, per-session VMs, private networking and RBAC, SSO, SCIM and OIDC support.

Health-data dashboards

Cognition's non-coding showcase features Bryan Johnson using Bots to ingest what the company describes as billions of data points into a synthesis of his health. The demo includes a dashboard spanning longevity, sleep and heart metrics.

Further reading

Discussion across the web

Where this story is being discussed, in original context.

On X· 7 threads
TL;DR2 posts
Persistent memory and child sessions2 posts
Ongoing goals and automations1 post
Needs You and bot-to-bot handoffs1 post
Android app in 44 days1 post
Audio regression testing1 post
Custom Bots and Unity testing2 posts
Share on X