Skip to content

Latest commit

 

History

History
35 lines (32 loc) · 2.06 KB

File metadata and controls

35 lines (32 loc) · 2.06 KB
title Devin - Long-running session agent
tagline End-to-end autonomous SDE in a cloud VM with browser, shell, and editor; one chat-style session runs hours-to-days unattended.
attribution Cognition Labs
tier tool
autonomous true
canonical_url https://devin.ai/
requires Devin subscription (≥ $500 / mo)
setup_command # In the Devin web UI or Slack integration: # State the goal once, in a single prompt: "Build feature X end-to-end, branch, code, tests, PR. Ping me on red CI or blocker." # Walk away. Review the resulting PR.
when_to_use Long async work where you'd otherwise context-switch repeatedly, predictable bug fixes, dependency upgrades, well-specified features.
when_not_to_use Tasks that require live debugging or per-step human judgment, Devin runs opaque mid-session.
tags
devin
cognition
async
session-agent
use_cases
building-features
sources
title author url year
Devin (Cognition Labs)
Cognition Labs
2024
related
antiPatterns practices tools workflows
cookie-clicker-mode-passive-supervision
yolo-mode-in-production
sandbox-for-autonomous-runs
human-in-loop-checkpoints
devin
cursor-background-agents
codex-cloud

Cognition shipped Devin in March 2024 and it drew equal parts awe and debunking videos. The product that survived that cycle is narrower and more useful than the launch demo: dependency upgrades, predictable bug fixes, well-specified features that nobody on the team wants to context-switch into.

The catch is opacity. Mid-session, Devin is a black box that occasionally asks questions in Slack. Everything you would normally steer interactively has to live in the brief: acceptance criteria, test commands, what "done" means. Cognition's own onboarding advice points the same direction: treat it like a new junior engineer and hand it small, well-specified tasks, with the PR review carrying the full weight of the review gate.

Sessions are metered in agent compute units (ACUs), so an underspecified task that wanders for hours costs real money on top of the review time.