AI Trends

Cognition Launches Devin Bots: Engineering Tool Supporting Multi-Agent Collaboration and Continuous Accountability

Shao MengUpdated Nov 18 18:29

AI summary

Cognition releases Devin Bots, allowing bots to have independent identities, memories, and tool permissions, breaking down objectives and launching multiple Devin sessions to work in parallel. Internal testing data shows that bots can automatically merge numerous pull requests and reduce bottlenecks, while humans primarily handle critical judgments and approvals.

Why read this

SelectedAI score 79

Devin Bots introduce continuous memory and multi-agent collaboration mechanisms, demonstrating their parallel processing capabilities in long-term engineering maintenance through real project data.

Topics

Tags

Quoted post

Cognition

@cognition

Meet Devin Bots. From writing firmware to merging thousands of PRs a week, a Devin Bot leads a team of agents to get the job done.

View original

Meet Devin Bots. From writing firmware to merging thousands of PRs a week, a Devin Bot leads a team of agents to get the job done.

View original post

Body · AI translation

Cognition Launches Devin Bots: From Taking Orders to Writing Code, to Leading Multiple Agents in Continuously Managing an Engineering Goal

Devin Bots aim to solve more than just "whether the Agent can complete this task"; they address who discovers the next step, checks results, and continues progress after a task ends.

Unlike Devin sessions, which correspond one-to-one with requests and executions, Bots are persistent engineering teammates: possessing independent identities, memory, and tool permissions, they can break down goals, launch multiple Devin sessions for parallel work, and then follow up on testing, code reviews, and subsequent improvements.

# What You Delegate Is Not a Ticket, But a Responsibility

An intuitive difference: instead of asking to "port this feature to Android," you ask the Bot to "continuously maintain feature parity between Android and iOS"; instead of "fixing this layout jump," you ask it to "continuously maintain page smoothness as the product evolves."

The former delivers once and ends; the latter requires continuously identifying gaps, scheduling work, and verifying results. Devin Bots incorporate this layer of continuous coordination into the product.

Bots can run on schedules or be triggered by events in tools, such as monitoring customer issues in Slack, checking CI nightly, or following up on blockers in Linear. When created, they can start from templates like QA, product development, or performance optimization, before configuring specific responsibilities.

# Bots Handle Coordination, Devin Sessions Handle Execution

Each Bot has a continuous conversation that receives goals, feedback, and progress updates. The specific tasks it breaks down are executed via Devin sub-sessions launched on independent cloud machines, operating in the same way as when an engineer directly starts a Devin session.

There are two types of long-term information here: one is human-set instructions that define responsibilities and working methods; the other is memory accumulated by the Bot during work, including preferences, corrections, and project details. Once shared with the team, everyone can collaborate around the same Bot without needing to re-explain context each time.

Bots can also delegate tasks to each other. In official demonstrations, a QA Bot identifies an issue and hands it off to a Developer Bot for fixing, then returns to verify. Humans do not need to manually move context through every step but instead focus on goal adjustments and key decisions.

# An Android App Case Study Demonstrates the Meaning of Continuous Responsibility

Cognition’s Android Bot built the Devin Android app in 44 days without dedicated Android engineers, merging 1,134 PRs, closing 1,217 tickets, and writing approximately 345,000 lines of Kotlin.

During this period, the person in charge sent only 183 messages—roughly one message for every six merged PRs—with 22 days where no messages were sent at all. The app has entered open testing on Google Play, and the Bot continues to follow up on features and test feedback.

Another Layout Bot builds the latest main branch code every three hours, finds, and fixes layout jumps. Within three weeks, it merged 159 related PRs, and the team reported a 70% reduction in stuttering. This illustrates the product direction better than single-session demos: responsibilities do not disappear when a session ends.

# Beyond Efficiency Numbers, What Matters More Is How Humans Intervene

Cognition stated that beta testers achieved an average of 4.5 times more merged PRs per message compared to previous levels. Jenny Xu ran a game studio using about 20 collaborating Bots, primarily communicating with a Super Jenny Bot responsible for overall coordination, checking in four times a day via mobile phone. Weekly PR volume increased from about 50 to 1,900, while maintaining weekly releases.

These numbers reflect changes in task parallelism and coordination methods, but PR volume does not directly equate to software value; it must be evaluated alongside feature delivery, quality, and maintenance costs.

Items requiring approval, option selection, or review are consolidated in "Needs You." Bots use their own configured connections and keys; they do not automatically inherit another person's permissions just because someone sends them a message. Permissions for usage, configuration, and sharing can be distinguished when shared. New access rights or automations require prior consent.

The product is now available for use, with usage counted against the plan of the creator; Teams and Enterprise plans can set spending limits or calculate usage caps.

Source

Shao Meng