⚡ Flash Agents

40 one-shot AI agents on Cerebras Inference — sub-second UX, zero chat overhead.

Paste → Run → Structured JSON. One call. No agent loops.

Quick Start · Example · Benchmarks · Architecture · vs LangChain

---

📋 Example input / output

Agent: Flash Review (GLM 4.7) · Latency: ~2.5s end-to-end (see benchmarks)

Input

function divide(a, b) {
  return a / b;
}
// no zero check

Output

{
  "summary": "Division sans garde — risque de crash runtime.",
  "score": 62,
  "critical": [
    {
      "line": "divide",
      "issue": "Division par zéro non gérée",
      "fix": "if (b === 0) throw new Error('Division by zero')"
    }
  ],
  "warnings": [],
  "positives": ["Fonction concise, lisible"],
  "verdict": "fix_first"
}

Same pattern for all 40 agents: one input → one JSON object → done.

---

🎯 What is this?

Flash Agents is an open-source showcase for ultra-fast LLM inference on Cerebras. Each agent does exactly one job with the best model for that job:

Model	Use case	Agents
GLM 4.7 (zai-glm-4.7)	Code, debug, regex, secrets	13
GPT OSS 120B (gpt-oss-120b)	Reasoning, product, career, security	20
Gemma 4 31B (gemma-4-31b)	Vision — screenshots, OCR, UI audit	7

Built for developers who want instant task completion without building a chatbot.

📊 Benchmarks

Measured on Flash Review (zai-glm-4.7), local Mac, Cerebras free tier, 5 runs with 15s rate-gate:

Metric	Flash Agents	Typical multi-agent chain
P50	2.6s	15–45s (plan → tool → synthesize)
P95	3.6s	30–90s
API calls	1	5–50+
Output	Valid JSON	Markdown + parsing needed

Run 1: 3621 ms  (cold)
Run 2: 2610 ms
Run 3: 2717 ms
Run 4: 2388 ms  ← P50 neighborhood
Run 5: 2543 ms

Reproduce: ./scripts/benchmark.sh (requires running backend + CEREBRAS_API_KEY).

Why not 300ms? Cerebras inference is fast; end-to-end includes network + JSON generation. The win is 1 call vs 10 — not raw token speed alone.

🏗 Architecture

flowchart LR
    A[Request] --> B[FastAPI Router]
    B --> C[Agent Registry]
    C --> D[Rate Gate]
    D --> E[Cerebras API]
    E --> F[JSON Response]
    F --> G[Guide + Latency ms]

    style A fill:#1a1a24,stroke:#2a2a38,color:#e8e8f0
    style E fill:#ff6b2c22,stroke:#ff6b2c,color:#ff6b2c
    style F fill:#3dd68c22,stroke:#3dd68c,color:#3dd68c

Loading

Layer	Role
Request	Text or image (vision agents) from React UI
Router	POST /api/run → picks agent by agent_id
Agent	System prompt + model + JSON schema
Cerebras	OpenAI-compatible inference (GLM / 120B / Gemma)
JSON	Parsed result + how_to + next_steps + latency_ms

✨ Highlights

• ⚡ One task = one API call — no multi-turn chat, no agent loops
• 📊 Structured JSON output — response_format: json_object, ready to pipe into tools
• 🖼️ Vision agents — upload PNG/JPEG, get UX audits, OCR, diagram explanations
• 🛡️ Rate-safe — built-in 15s gate for Cerebras free tier (~5 req/min)
• 🔐 Local secrets — API key in .env on your machine, never sent to the browser
• 🧭 Onboarding built-in — explains the app + validates your Cerebras key on first run
• 🍎 macOS launcher — double-click .command on Desktop to start everything

🚀 Quick Start

1. Get a free Cerebras API key

→ cloud.cerebras.ai (models: zai-glm-4.7, gpt-oss-120b, gemma-4-31b)

2. Install & run

git clone https://github.com/anthonyoccelli33480-ctrl/flash-agents.git
cd flash-agents
cp .env.example .env   # add CEREBRAS_API_KEY=csk-...

make install           # Python 3.12 venv + npm
make backend           # http://127.0.0.1:8787
make frontend          # http://127.0.0.1:5173

Or on macOS: double-click ⚡ Lancer Flash Agents.command from the Desktop folder.

3. Try an agent

Try this	Input	Output
Flash Review	Paste a code diff	Score, critical issues, verdict
Flash Fix	Stack trace + code	Root cause + patch
Flash UI Review	Screenshot PNG	UX critique JSON
Flash MVP	Vague startup idea	Scoped MVP plan

🤖 Agent Catalog

Dev (11) — GLM 4.7

Agent	One-liner
Flash Review	Code review → structured score + fixes
Flash Fix	Stack trace → root cause + patch
Flash Commit	Git diff → Conventional Commits + PR
Flash Regex	Natural language → tested regex
Flash Test	Function → unit tests
Flash SQL	Question + schema → SQL query
Flash Explain	Code → clear explanation
Flash Refactor	Code smell → clean version
Flash Docker	Stack → minimal Dockerfile
Flash OpenAPI	Endpoints → OpenAPI spec
Flash Migrate	Framework A → migration plan B

Career (5) — 120B + README on GLM

Flash JD · Flash STAR · Flash LinkedIn · Flash Email · Flash README

Product (6) — GPT OSS 120B

Flash MVP · Flash Decision · Flash Pitch · Flash Competitor · Flash Pricing · Flash Onboarding

AI / ML (5) — 120B + Router on GLM

Flash Eval · Flash Prompt · Flash Dataset · Flash Rubric · Flash Router

Content (3) — GPT OSS 120B

Flash TL;DR · Flash Compare · Flash Outline

Security (3) — 120B + Secret on GLM

Flash Threat · Flash Secret · Flash RGPD

Vision (7) — Gemma 4 31B + image upload

Flash UI Review · Flash Diagram · Flash Wireframe · Flash OCR · Flash A11y · Flash Chart · Flash Mockup

Full reference → docs/AGENTS.md

🆚 Why Flash Agents instead of LangChain / CrewAI / OpenAI Agents?

	Flash Agents	LangChain	CrewAI	OpenAI Agents SDK
Mental model	One-shot task	Chains & graphs	Multi-agent crew	Tool loop
Calls per task	1	3–20+	10–50+	5–30+
Output format	Forced JSON	Usually text	Markdown reports	Tool results + text
Latency (typical)	2–4s	10–60s	30–120s	5–40s
Infra required	FastAPI + React	LangSmith, vector DB…	Roles, tasks, memory	OpenAI + tools setup
Inference	Cerebras (GLM, 120B, Gemma)	Provider-agnostic	Provider-agnostic	OpenAI only
Best for	Dev tools, quick JSON tasks	Complex pipelines	Autonomous teams	OpenAI ecosystem
Learning curve	Clone & run	Steep	Medium	Medium

Use Flash Agents when: you want a code review, SQL query, MVP scope, or UI audit in one click — not a research project.

Use LangChain/CrewAI when: you need multi-step autonomy, RAG, or agents that collaborate over minutes.

Complement, not competitor — Flash Agents is what Cerebras feels like without the framework tax.

➕ Add an agent

Edit backend/app/agents/registry.py:

_register(AgentDef(
    id="my_agent",
    name="Flash My Agent",
    tagline="Input → output in one shot",
    model="zai-glm-4.7",  # or gpt-oss-120b, gemma-4-31b
    category="dev",
    icon="⚡",
    placeholder="Paste your input…",
    system="""Respond ONLY in JSON: { "result": "..." }""",
))

🤖 Machine-readable

Agents, API, stack, and ecosystem for LLM crawlers and tooling:

• PROJECT.json — structured project manifest
• metadata.jsonld — Schema.org SoftwareApplication

🔑 Keywords

cerebras · llm-agents · ai-agents · inference · fastapi · react · vision-llm · gemma · glm-4 · code-review · developer-tools · one-shot-agents · structured-output · open-source

📜 License

MIT — see LICENSE

---

If Flash Agents saved you time, a ⭐ helps others find it.

Built to show what Cerebras Inference feels like when you don't slow it down.