Parallel agents are a false good idea
Parallel agents felt like rolling dice. I could not follow, could not answer, then had to test a feature bigger than I would have shipped. Going local forced a slower, clearer rhythm.
Parallel agents are several agents running at once, each on its worktree. On paper that looks like huge productivity. In practice I could not keep up, I lost the thread, and I had to test a feature bigger than I would normally have shipped.
This sits next to the perpetual application map: one local agent that you can actually follow, not a swarm you cannot brief.
Canonical URL: https://www.zheat.xyz/en/insights/parallel-agents-false-good-idea/
Table of contents
- What are parallel agents?
- Why do they feel like rolling dice?
- What changed with 100% local?
- Parallel agents vs a local agent
- FAQ
- Related
What are parallel agents?
Parallel agents are several agents launched at the same time, each working on its own worktree.
On paper it is tempting. Huge productivity. What I actually noticed: I cannot keep up. Worse, sometimes I no longer know where I am. When an agent asks me a question, I struggle to answer.
Why do they feel like rolling dice?
The more agents I launch in parallel, the more it feels like rolling dice. By trying to be everywhere, I end up really nowhere.
At the end I have to check and test a feature much bigger than what I would normally have produced. The score is not that positive.
What changed with 100% local?
I moved to a 100% local approach, so my capacity to parallelise agents is much lower anyway. That may be the best thing that could have happened.
In the coming months I will keep testing 100% local, at a slightly less frantic pace. Early feedback is very satisfying. You have to pay more attention, and be more precise about intent. At least I am no longer drowned by the progress of my own product.
Produce less in parallel. Keep control. Keep your head.
Parallel agents vs a local agent
| Parallel agents | 100% local | |
|---|---|---|
| Setup | Several agents, each on a worktree | Capacity to run many at once is much lower |
| Follow-through | Cannot keep up. Lose the thread. Struggle to answer when an agent asks. | Slower, clearer rhythm. More attention. More precise intent. |
| Review | Test a feature bigger than you would have shipped | Produce less in parallel, keep control |
| Fit with a perpetual app | A swarm you cannot brief | One local agent you can follow |
FAQ
Are parallel agents a good idea? They look good on paper. In this setup they felt like rolling dice: I could not follow, I could not answer, then I had to test a feature bigger than I would have shipped.
What is a parallel agent setup? Several agents launched at once, each on its worktree.
What does 100% local change? Capacity to parallelise drops. You pay more attention and you have to be more precise about intent. You are no longer drowned by the progress of your own product.
How does this relate to a perpetual application? A perpetual application wants one local agent you can actually follow, not a swarm you cannot brief. See the perpetual application map.