LLenny's Podcast
← All frameworks
ProductivityScott Wu (CEO and co-founder of Cognition)

The Agent Foothold Onboarding

Onboard an AI agent like a new hire: environment first, easy tasks next, then scale.

Difficulty
Easy
Time to result
~days to results
Steps
3
Confidence
85%

Autonomous agents fail when dropped cold onto big work. Cognition's onboarding pattern treats the agent like a junior engineer joining the team — set up its environment and access, give it small 'one-pointer' tasks to build a foothold and codebase familiarity, then progressively scale up scope as it accumulates knowledge.

Origin

Scott Wu, from lessons learned watching users struggle with Devin's blank-screen start.

Core principles

  • 01An agent needs a working environment before it can be productive
  • 02Small first tasks build codebase familiarity and a foundation to grow from
  • 03Capability comes from accumulated context, so knowledge compounds with use
  • 04The agent should be able to run and verify its own work (tests, lint, CI)

How to run it

  1. 1

    Set up the repository and environment

    Get the repo connected and ensure the agent has what it needs to actually work — its own virtual machine, and the ability to run tests, the linter, and CI.

    Pro tip Wire in integrations (Slack, GitHub, issue tracker) so handoffs happen where work already lives.

    Watch out Skipping environment setup means the agent can't verify its own code, so its output can't be trusted.

  2. 2

    Hand it one-pointer tasks

    Start with a couple of small, easy tasks so the agent gets familiar with the codebase and builds an initial foothold before tackling anything large.

    Pro tip Pick tasks with obvious verification so early wins are unambiguous.

    Watch out Opening with a 'rearchitect everything' request wastes the onboarding and produces poor results.

  3. 3

    Scale scope over time

    As the agent proves itself and accumulates codebase knowledge, progressively hand it bigger and more central work, learning how best to communicate and which projects to bring it in on.

    Pro tip Let the agent build a wiki/representation of the codebase — humans can use it for onboarding too.

In the wild

From blank screen to knocking out PRs

Users who set up the repo, taught Devin to run lint and CI, and started with initial foothold tasks found it soon 'writing all these PRs' like a new engineer who'd just joined.

A properly onboarded agent moves from confused blank-screen start to reliably shipping pull requests.

Common mistakes

Treating it as instant magic

Expecting productivity without environment setup or foothold tasks leads to the common early failure of asking for a huge rearchitecture on day one.

Is it for you?

Best for

Teams introducing an autonomous coding agent into an existing codebase for the first time

Not ideal for

Throwaway one-off prototypes where the setup investment outweighs the payoff

From the transcript

let's make sure we we hand Devon a couple onep pointer tasks so it can get familiar with the codebase

18:30

we want to make sure Devon's got its own virtual machine set up to be able to do that

18:30

if Devon needs to be able to test the code or run the llinter or CI

18:30

From the episode

Inside Devin: The world’s first autonomous AI engineer that's set to write 50% of its company’s code by end of year

Scott Wu (CEO and co-founder of Cognition)