Rodrigo Lemela Duarte

Founder da | Engenharia de software e IA aplicada

[Rodrigo Lemela Duarte @RodrigoLemela]

@0xAndrewShu each agent has its own machine and worktree, and every test run gets its own database schema, so parallel runs don't collide. the full suite runs one PR at a time on a single box before anything merges, and the agents queue that up in agent-chat.

[Rodrigo Lemela Duarte @RodrigoLemela]

@Quantradin @fabienpenso yes. a reservation lasts two minutes unless the agent keeps renewing it while it writes, and never more than ten. after that the reply is open again, and a person in the room can also release it. worth trying with your own agents.

[Rodrigo Lemela Duarte @RodrigoLemela]

@qwerty__cs Reading the previous session's log covers it when both agents run on the same machine. Across a team the next agent is usually on someone else's laptop, which is why we built agent-chat with the team chat in the cloud, no Mac, Tailscale or local database in the middle.

[Rodrigo Lemela Duarte @RodrigoLemela]

@ai_binbin Traces also catch the run that went sideways because the agent didn't know what another session had just changed. That never shows up when you scroll a single chat.

[Rodrigo Lemela Duarte @RodrigoLemela]

@Asteri_eth Merging orders and shipping because one ran right after the other in every trace is a clean rule. The same test probably works one level up, where two agents that always hand work to each other are really one loop.

[Rodrigo Lemela Duarte @RodrigoLemela]

@sermakarevich The turn 28 failure gets worse when a different agent picks up the ticket. It usually inherits the summary but not the reason behind a decision, so it redoes something the first one already ruled out.

[Rodrigo Lemela Duarte @RodrigoLemela]

@icanvardar That loop is still most of the real work. The swarm demos skip the part where someone says no every few minutes, and with more agents that person just gets busier.

[Rodrigo Lemela Duarte @RodrigoLemela]

@PurpleSuede22 a heroic amount of copy/paste is where we started too. what fixed it for us was one shared room in the cloud: Codex on one machine and Claude on another post their handoffs there, and each one gets pinged when someone mentions it. that's what agent-chat grew out of.

[Rodrigo Lemela Duarte @RodrigoLemela]

@Quantradin @fabienpenso both. claiming a card is all or nothing, so one agent wins and the other is told it's taken, by whom and when. the talk still happens in the room, and when a message mentions several agents, one reserves the reply first so they don't talk over each other.

[Rodrigo Lemela Duarte @RodrigoLemela]

@Kenny_GTM @Frontal_AI one repo keeps everyone on the same setup, which is half of it. the other half for us was seeing what each person's agent has done and is working on, so we built agent-chat, a team chat in the cloud where those agents post their work and get pinged when mentioned.

[Rodrigo Lemela Duarte @RodrigoLemela]

@r1ckdeoliveira agents talking to agents gets wild fast. the part we kept missing was where the humans fit in, so we built agent-chat, where people and their claude code or codex agents share the same rooms and see each other's messages.

[Rodrigo Lemela Duarte @RodrigoLemela]

A parte boa? Esperaram dois dias da eleição pra divulgar.
VOA BRASIL!

[O Analista @Woodyraiz]

Pelo visto, com o IBS, a alíquota total vai passar de 30% do faturamento da empresas.

Parabéns a todos os envolvidos.

[Rodrigo Lemela Duarte @RodrigoLemela]

Hahahahahahaha
Sem maiores comentários hahahah

[Expo @expo]

Native iPhone Duo apps, built with Expo.

Ship yours before your AppleCare appointment!

[Rodrigo Lemela Duarte @RodrigoLemela]

@stretchcloud Three in one hour says the pain is real. Most of them fix one dev with many agents, though. It gets harder when two people's agents need the same context.

[Rodrigo Lemela Duarte @RodrigoLemela]

@alimwtf We didn't, for a while. What helped was having the agents post their work where the team already talks, so each one sees what the others picked up.

[Rodrigo Lemela Duarte @RodrigoLemela]

@RoundtableSpace Passing context between agents is the real bottleneck once you run more than one. We ended up putting our agents and the team in the same chat, so the history is already there instead of something you dig up later.

[Rodrigo Lemela Duarte @RodrigoLemela]

@FFmpeg DM'd you guys.

[Rodrigo Lemela Duarte @RodrigoLemela]

@relentangleRS there's a free plan, up to 3 people and 3 agents turned on, no card needed. it's its own chat room, not slack, and you plug in grok bot and dots with your own accounts. https://t.co/1iORhSJG8P