LA Hacks 2026 · UCLA

You hire AI agents,
not write posts.

Onboard a brand once. AgentBuffer spawns a small team of AI agents tied to that brand. They wake up on a schedule, generate on-brand content, run it past a Critic that must reject weak work, then auto-publish. While you sleep.

AgentBuffer - Buffer for autonomous brand agents

Multi-agent, on purpose.

Three uAgents per brand on Fetch.ai's Agentverse. Each one has one job. The Critic is allowed to say no.

Hire, don't write

Onboard a brand once (Q&A, PDFs, past videos, linked socials). Cognition agents spawn from a "hire" form and post on the cadence you pick.

Critic must reject

An adversarial Critic uAgent grades every proposed slot on a 5-axis rubric. Rejection forces revision. No rubber-stamping.

Direct to six platforms

Publisher posts directly to LinkedIn, X, Instagram, TikTok, YouTube, and Bluesky via platform APIs. Idempotent, with replay.

Real cadence

Each cognition agent is a uAgent with on_interval(parse_cadence(row)). Deterministic Agentverse address, stable across restarts.

How it actually runs.

Three shared helpers, one cognition agent per brand. Critic loops back when work isn't good enough.

AgentBuffer architecture - user, orchestrators, cognition agent, strategist, critic, publisher, media, platforms

Built with.

Six sponsor stacks, each used for what it's actually best at.

Fetch.ai · Agentverse
3 uAgents (Strategist, Critic, Publisher) + per-brand cognition agents. ASI:One NL ranker.
Anthropic · Claude Sonnet 4.5
Brand-kit extraction from PDFs + Critic rubric. Visible failures, no 4.6/5 rubber-stamps.
Google · Nano Banana 2 + Veo 3.1
Same-seed image iteration when the Critic flags a crop. Veo at onboarding only.
Cognition · Devin
Owns Publisher reliability end-to-end: idempotency, backoff, dead-letters, replay CLI, fault injection.
Cloudinary
Server-side l_fetch brand-bar overlay with named transforms. No exact-hex Sharp passes.
Supabase + Vercel
Auth Hook for org-scoped JWT, RLS by org_id, Postgres + Storage, Next.js 15 deploy.

Built for LA Hacks 2026.

A 36-hour project by team AgentBuffer at UCLA. Demo-day climax: a Critic uAgent rejecting a real piece of work, live.

View on GitHub