Variant A · Signal Light · zero-animation build

Heterogeneous agent teams, orchestrated from one CLI.

Claude Code conducts; Kimi K3 and MiniMax-M3 workers execute under one task contract. Every claim on this page comes from a real run in this repository — no staged screenshots, no invented metrics.

See the capability demos Read the lab notes
100/100repo tests passing at audit time
4 agentsread-only swarm audit of core source & tests
6/6swarm citations spot-verified against source
−46%words when K3 condenses prose, caveman style · +11% on generation, so it stays normal

Four capabilities, each with a runnable demo

Each pattern is documented on its own static page and backed by a live aflow workflow you can re-run.

Review

Long-context code review

K3 reads the target and produces evidence-backed findings; MiniMax-M3 cross-checks for misses.

 Swarm

AgentSwarm exploration

Parallel read-only subagents inspect disjoint scopes; the lead synthesizes one verified answer.

 Resilience

Quota-aware fallback

402 / entitlement errors classify as quota, so chains degrade Kimi → MiniMax-M3 without human input.

 Prototyping

Prompt-to-site

Static page concepts generated as safe, reviewable artifacts inside the lab boundary.

The agentic loop, structural not theatrical

Gates are enforced by the DAG: downstream agents never spawn until the upstream gate clears. This is the same loop that produced this page.

Dispatch

One task contract fans out to heterogeneous workers — kimi, MiniMax, or claude profiles.

Review

K3 reviews the diff or files and returns schema-validated findings, never free-form prose.

Cross-check

A second vendor verifies each finding and hunts for what the first pass missed.

Gate

Severity-ranked results halt at gate: pass; a human approves before anything ships.

Run the backing workflow yourself

The visuals are static; the orchestration is real. Expected on the sample fixture: a halt at gate: pass, because the fixture ships a deliberate high-severity bug.

node bin/aflow.js workflow run top-review-lab --variant strict
# or the two-model corroboration loop:
aflow flow run crosscheck --diff HEAD~1 "review this change"

Safety boundaries

  • No credentials, transcripts, or session state in any demo page.
  • No mutation outside experiments/skills-workflows-hooks-lab/.
  • All pages static HTML/CSS — no backend, no trackers, no animation budget.