Aweb.ai runs 7 named persistent AI agents plus ephemeral coding agents with shared taskboards, durable artifacts, and async messaging – two humans set direction.
Key Takeaways
AI-native differs from AI-assisted: coordination happens between agents, not humans relaying work between each other.
Each agent gets a stable identity via directory path (e.g. ~/agents/athena/), role doc in AGENTS.md, and messaging via the open-source aw CLI.
Durable artifacts – tasks, decision records, handoffs, verified-live mails – are required; conversations evaporate, artifacts survive across sessions.
Specialization compounds: agents accumulate months of role-specific context that a fresh prompt cannot replicate.
A public template repo (github.com/awebai/agent-first-company-template) provides decision-record, handoff, and status-file structures to fork.
Hacker News Comment Review
The post landed almost entirely negative: commenters dismissed it as AI-generated content, with no engagement on the technical architecture or operational principles.
One commenter raised a legitimate coordination critique: AI produces output faster than humans can evaluate it, so agent-heavy workflows may worsen rather than solve the coordination problem.
