Peter Broas

run claude code, codex and cursor agents side by side on one canvas. windows + mac. building it in public, something new every day ⚡

[Peter Broas @DashCanvas]

Introducing DASH.

Every AI agent on one canvas. Give each one a job, watch them work side by side, and see the changes land live.

Free to try on Windows & macOS 👇

[Peter Broas @DashCanvas]

@darkgaro when a session moves to another account mid task, does the agent keep its context or restart from the transcript

[Peter Broas @DashCanvas]

@0xinhua which model and effort level? haiku wrote 3.5x to 5x the tokens of the others on the same bug fix for me, so overbuilding varies a lot by model

[Peter Broas @DashCanvas]

@stretchcloud green tests aren't the same as the same fix. in a 4 agent bug hunt all passed 1,699 tests but two of them fixed one bug differently than the original

[Peter Broas @DashCanvas]

@DavidSZD1 how many runs per model? same fable medium took me 3:04 once and 2:05 another time on one 6 bug fix

[Peter Broas @DashCanvas]

@Voxyz_ai how do you check opus didn't drop something from the plan fable wrote, do you diff the two or just run the tests

[Peter Broas @DashCanvas]

@fawnunwrap @coah80 in a 6 bug fix i ran, fable medium beat opus high on time 3 runs of 3 (2:20 vs 2:42 median) but wrote about 10% more output tokens

[Peter Broas @DashCanvas]

each run: fresh session, its own copy of qs with fresh git history (no diffing to the answer), claude code's default effort (opus runs high by default, the other three medium).
per run: fable 3:04 2:05 2:20 · opus 3:42 2:33 2:42 · sonnet 3:16 3:51 3:01 · haiku 4:35 3:48 4:08
fable was fastest but not the leanest: 3,509 avg output tokens vs opus 3,189.
try it 3 days free: https://t.co/afXChEjyeH

[Peter Broas @DashCanvas]

fable 5.2 on medium beat opus 5.5 on high, 3 runs out of 3 (both claude code defaults)

same prompt, 4 claude code agents side by side in dash, 6 bugs i planted in qs. median time to green:

fable 5.2: 2:20
opus 5.5: 2:42
sonnet 5.5: 3:16
haiku 5.5: 4:08

12 of 12 runs fixed all six. haiku wrote 3.5x to 5x the tokens of the others in every run

[Peter Broas @DashCanvas]

@kloss_xyz how much of the 17 minutes was rendering and how much was writing it

[Peter Broas @DashCanvas]

@johnennis when it coordinates, is it handing work to other agents or running every step itself

[Peter Broas @DashCanvas]

@9ranty @TellaHQ same here. a 28 second 1080p cut takes my laptop about 9 minutes to render, longer than the agents took to build what's in it

[Peter Broas @DashCanvas]

@Signal_65 small data point: fable 5.2 won all 3 runs of a 6 bug race against opus, sonnet and haiku in claude code. median 2:20 vs opus 2:42

[Peter Broas @DashCanvas]

@rohit3a matches a test i ran today: 6 planted bugs, 3 runs each. median opus 2:42, sonnet 3:16, fable 2:20. opus and sonnet output tokens within 7% of each other

[Peter Broas @DashCanvas]

@be1sidd @ClaudeDevs same prompt last night, fable 5.2 finished in 2m 16s and opus 5.5 took 4m 57s with twice the code

[Peter Broas @DashCanvas]

@CodeTiwg what kind of task finally made it fork, if any

[Peter Broas @DashCanvas]

@kweinmeister what fixed it for us was file ownership. one agent literally said "that file is ben's, so i left it alone"

[Peter Broas @DashCanvas]

@choopyplug1 which of the 10 saved the most after the subagent model one

[Peter Broas @DashCanvas]

@talperetz_ worth putting .claude/settings.local.json in your global gitignore, not just per repo

[Peter Broas @DashCanvas]

@rileybrown did it get the per channel role permissions right too, or just the server level ones

[Peter Broas @DashCanvas]

@rodri_shema every agent's window on one canvas with a live git diff next to them, and they move their own task board cards. it's why i built dash