The Parallel Agent Workflow
Run several AI agents at once and insert yourself only where your expertise actually matters.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 4
- Confidence
- 90%
Rather than working synchronously on one task, Cognition's engineers each run up to five agents in parallel, kicking off multiple pieces of work asynchronously and stepping in only to define the problem or steer complex parts. This shifts the human from single-threaded executor to a director of parallel work.
Origin
Scott Wu describing how Cognition's ~15-person engineering team operates, with roughly five Devins per engineer.
Core principles
- 01Humans are single-threaded on context; agents let you run multiple lines of work at once
- 02Your leverage is the parts that need your expertise, not the bulk implementation
- 03Asynchronous handoff, not synchronous babysitting, is where the multiplier comes from
- 04It takes real practice before parallel operation feels intuitive
How to run it
- 1
List the day's discrete tasks
Decide the handful of things (roughly five) you want done today and break them into separate, parallelizable units of work.
Pro tip Bugs, feature requests, tests, and migrations are natural parallel candidates.
- 2
Assign one agent per task
Spin up a separate agent for each task and kick them all off asynchronously — agent one does task one, agent two task two, and so on.
Pro tip You can start and shut down agents as you see fit; treat capacity as elastic.
- 3
Stay for the expertise-critical moments
Jump in only where it counts: to precisely define what you're solving for, or to steer the more complex changes (e.g. how a class should be structured and its downstream references updated).
Pro tip You can answer an agent's clarifying questions at any point and hand back off — it won't lose the thread.
Watch out Trying to watch every agent action-by-action defeats the parallelism; only jump in as needed.
- 4
Review, merge, and capture learning
Come in at the end to review each output, merge what's good, give short feedback on the rest, and let the agents accumulate knowledge of your codebase and process over time.
In the wild
Cognition's team of about 15 engineers each run up to five Devins, which together merge several hundred pull requests into production every month — around a quarter of all code shipped, projected to exceed half by year-end.
→ A small team ships a sizable fraction of its production code through parallel asynchronous agents.
Common mistakes
Working the agents synchronously
Sitting and watching one agent step by step reproduces single-threaded work and forfeits the whole point of parallelism.
Expecting it to feel natural immediately
Wu notes it took the team time to adjust; treating the first awkward week as failure causes people to abandon the workflow too early.
Is it for you?
Best for
Engineers and small teams wanting to multiply output with autonomous agents already integrated into their stack
Not ideal for
Single deeply-coupled tasks that can't be decomposed, or teams not yet set up to verify agent output quickly
From the transcript
“working with with up to five Devons at once”
“what are the five things that we want to get done today”
“the goal for each of your tasks is to be there for the parts that really need your expertise”
“Devon merges like several hundred pull requests into production in the Devon code bases every month”
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)