@amasad describes Replit's agent loop architecture
@amasad describes a "loops" workflow at Replit where an orchestrator drives parallel agents, a computer-use verifier provides feedback, and specialized security, production, and SEO agents generate fix prompts — with minimal human prompting.
Score breakdown
The post describes a working multi-agent loop architecture where human prompting is reduced to a single outcome sentence and specialized agents handle orchestration, verification, and fix generation autonomously — a concrete example of minimal-human-input agentic coding in production at Replit.
- 01@amasad uses a workflow he calls 'loops' instead of traditional prompting
- 02Most human prompts in this workflow are barely a sentence, expressing only an outcome
- 03An orchestrator agent spawns and manages parallel agents
In a post on X, @amasad outlines the agentic architecture he has been using at Replit, which he calls "loops." Rather than crafting detailed prompts, he says most of his inputs are barely a sentence long — just enough to express an intended outcome. The heavy lifting is handled by a layered agent system: an orchestrator that runs parallel agents, a computer-use verifier that provides feedback, and specialized agents for security, production, and SEO that autonomously generate prompts for fixes.
@amasad positions this as a leading-edge practice, asserting that the broader industry is typically 3–6 months behind what Replit is doing with this loop-based approach.
Key facts
- 01@amasad uses a workflow he calls 'loops' instead of traditional prompting
- 02Most human prompts in this workflow are barely a sentence, expressing only an outcome
- 03An orchestrator agent spawns and manages parallel agents
- 04A computer-use verifier agent provides feedback to the loop
- 05Specialized security, production, and SEO agents generate their own fix prompts
- 06@amasad claims the industry is typically 3–6 months behind Replit's practices
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 13, 2026 · 08:58 UTC. How this works →