Adam Sip

One person, many agents. I post the unglamorous part: where work hands off between them, and where it drops. Board they pull from:

[Adam Sip @AdamSip_]

One person, many agents.

The hard part isn't the model — it's handoff. One finishes; the next needs state, not my chat log.

So work sits on a board they pull from, cards come back with evidence.

https://t.co/IW0P4natrl

[Adam Sip @AdamSip_]

@AlexFinn keyword: API 😂 gl with that.

[Adam Sip @AdamSip_]

@rhowardstone @karolzdeb @alekgrygier i guess mostly by securing copyrights... yes that's tough one... who owns rights to ai generated content 😅 especially with bad actors who slop generate enormous amount of music, clips, faces, etc. from my point of view art become commodity, marketing and distribution - asset 😜

[Adam Sip @AdamSip_]

@rhowardstone @karolzdeb @alekgrygier Artists must adapt to the changing technology. It was always like that, but ai put bar so low that competition will be enormous. it's revolution.

[Adam Sip @AdamSip_]

@vinceservidad same here ;) hopefully they give full reset soon

[Adam Sip @AdamSip_]

@JohnLrDev Astra max for CEO, Astra med for the Control Room, sol med/high for the rest - 50% of the weekly limit spent in 1 day 😅 need to compare/adjust

[Adam Sip @AdamSip_]

AI agents don’t just hallucinate facts. They hallucinate architecture. Ask for a simple content loop and they may invent schedulers, ledgers, recovery systems and abstractions. A useful safeguard: force the smallest testable path first. Add machinery only after it fails.

[Adam Sip @AdamSip_]

Practical tip: if you use Ponytail, most of this is already covered.
But add one more rule to your agent instructions: “A test must not be more complex than the behavior it protects.”

[AI Builder Club @aibuilderclub_]

How to stop GPT-5.6 Sol from overengineering your codebase.

Sol-class models don't write wrong code. They write too much code.

You ask for a config change, you get a framework. You fix a bug, you get an adapter layer.

Tests pass, CI green, looks professional. Then you touch one thing next week and there are four layers between you and the logic.

Fix below.👇

[Adam Sip @AdamSip_]

@jbarbier @YoussefCharifH2 Worktrees solve the branch collision. Have you also separated write authority? I’ve found isolated sessions can still conflict unless one owns each shared side effect and the others stay read-only until ownership is explicit.

[Adam Sip @AdamSip_]

An agent handoff can be saved and still fail.

"Saved" proves the artifact exists. It does not prove anyone received it.

Track three states:
saved -> delivered -> acknowledged

No acknowledgment? The handoff is still open.

[Adam Sip @AdamSip_]

@JohnLrDev Quand tu dis « assignées toutes seules », qu’est-ce qui déclenche l’assignation — une GitHub Action ou l’agent qui claim l’issue ? C’est précisément le maillon que j’essaie de fiabiliser chez moi.

[Adam Sip @AdamSip_]

@jskalc @orca_build With Pi, Claude, GPT on separate worktrees, who claims the next piece — you, or an idle agent?

[Adam Sip @AdamSip_]

@dMacro_dBS When Codex finishes implement and Claude finishes review, who picks the next task — you, or does an idle session pull it?

[Adam Sip @AdamSip_]

@JohnLrDev mac mini m4 here
- chatgpt desktop for "jarvis" (orkchestrator) with voice live (as "control room")
- cmux with all the CLIs from different providers/harnesess (jarvis use relay to assign tasks to them
- grok bot desktop+cli in cmux (experimental atm)

[Adam Sip @AdamSip_]

@JohnLrDev Chez moi la suivante n'est pas prise par un agent idle — elle doit déjà être sur un tableau et réservée. L'agent ne se réveille pas tout seul. Chez toi les Issues s'assignent toutes seules, ou tu cliques encore ?

[Adam Sip @AdamSip_]

@B0SS_8888 @thsottiaux When the plan lives in chat and Codex starts empty, who still owns the handoff — you paste, or does something already hold the plan?

[Adam Sip @AdamSip_]

@bygregorr @TonyXavier_ When both queue edits on the same file, who wins — worktree per agent, or does something claim the path before either writes?

[Adam Sip @AdamSip_]

@LongCandles When pull-first already killed the copy-paste, what still has to stay on you before an agent is allowed to grab the next review itself?

[Adam Sip @AdamSip_]

@mhpaler File claim + one bead per agent stops collisions in one repo. When beads sit on different tracks (build vs review vs ship), who pulls the next one — idle agent, or do you still assign?