Pedro Arantes

@arantespp

Building stuffs using AI 🌳 naturali.ai 🌊 flow.ttoss.dev 🪙 myhold.app 🎯 oneclickads.com.br

Born 1993

Joined December 2012

380 Following

2,547 Followers

Replies Reposts Media Articles

Pedro Arantes
@arantespp
15h
Faster agents won’t speed up delivery if the time just turns into waiting. Think of a task as a trip through several stations: write the code, review it, approve it, deploy it. Delivery time is the sum of every station plus the time spent waiting between them. Agents make one
1
Pedro Arantes
@arantespp
17h
Agents write code faster than people can review it. So the cost of a project moves from doing the work to the queues around it: the review waiting on you, the decision nobody made, the PR that came back. Today I shipped a change to Flow MCP so a team can see those queues. The
Pedro Arantes
@arantespp
Nov 1
Some definitions of my agentic workflow.
Pedro Arantes
@arantespp
Nov 1
Replying to @arantespp
Some definitions: Graph: the record of all my work, across every repo. Tasks, what blocks what, the PRs, the assumptions, the decisions. It’s the only memory the system has. Task: one unit of work in the graph. It has a status, a holder, and links to whatever it depends on.
Pedro Arantes
@arantespp
Nov 1
Back to one line from my first post: parallel agents never act on stale state. People asked how that works in practice, so here it is. The problem: all my agents act as me. Same account, same credentials. The graph’s “assignee” field can’t tell them apart, so “assigned to
1
Pedro Arantes
@arantespp
Oct 31
Third principle of my agent workflow: report, don’t refuse. Agents keep the graph updated through small writes nobody asked for: a status here, a dependency there. But when one of those writes fails unexpectedly, the agent learns that writing is risky. It starts double-checking,