Lár: The PyTorch
for Agents.

Stop guessing with black boxes. Start engineering with auditable, deterministic AI workflows.

Get Started Try Architect Beta

Stop Guessing. Start Knowing.

We built the same agent with LangChain and Lár.
When we hit a 429 Rate Limit error, this is what happened.

The Black Box

Opaque failures. Infinite retries. 100-line stack traces.

Traceback (most recent call last):

File "support_app_lc.py", line 144, in <module>

final_response = full_agent_chain.invoke({"task": task})

File "langchain_core/runnables/base.py", line 2877

input = context.run(step.invoke, input, config, **kwargs)

File "tenacity/__init__.py", line 418, in exc_check

raise retry_exc.reraise()

File "concurrent/futures/_base.py", line 449

return self.__get_result()

File "google/api_core/grpc_helpers.py", line 77

raise exceptions.from_grpc_error(exc) from exc

google.api_core.exceptions.ResourceExhausted: 429 You exceeded your current quota, please check your plan and billing details.

The Lár Glass Box

Structured errors. Precise status codes. Instant clarity.

{
  "step": 4,
  "node": "LLMNode",
  "error": {
    "code": 429,
    "message": "You exceeded your current quota...",
    "status": "RESOURCE_EXHAUSTED",
    "details": [
      {
        "@type": "type.googleapis.com/google.rpc.QuotaFailure",
        "violations": [
          {
            "quotaMetric": "generate_content_free_tier_requests",
            "quotaValue": "2"
          }
        ]
      },
      {
        "@type": "type.googleapis.com/google.rpc.RetryInfo",
        "retryDelay": "9s"
      }
    ]
  }
}

Observability is Built In.

Why pay for LangSmith when Lár does it for free?
Lár stops execution the moment an error hits—saving your tokens—and delivers a pristine, replayable log of the exact failure state.

Zero Guesswork Deterministic Replays

Zero Waste Stop on Error

100% Open Source Apache 2.0

Lár: The First GxP-Ready Agent Framework.

Deterministic. Auditable. Air-Gap Capable.

Lár generates 21 CFR Part 11-Ready audit trails out of the box. Stop building chatbots. Start building regulated workflows.

Complete Audit Trails

Every step is logged in a full, immutable flight log. Essential for:

FDA 21 CFR Part 11 (Electronic Records)
HIPAA Audit Requirements
Clinical Trial Documentation (ICH-GCP)
Medical Device Regulations (ISO 13485)

Scientific Method (Not Chatbots)

Scientists hate chatbots that hallucinate. They need reproducibility.

Guaranteed Reproducibility: Same seed = Same path.
Forensic Logs: Debug decision trees, not random text.
Research Grade: Bring the Scientific Method to AI.

Air-Gap Capable (SCIF Ready)

Secure biomedical research often happens offline.

Build unclassified on laptop.
Serialize to JSON artifact.
Carry to SCIF on thumb drive.
Run locally with Snath Enterprise.

Granular Cost Tracking

Track token usage per node, not just per run.

Vital for budget compliance in research grants and healthcare economics where every inference cost must be accounted for.

"Is this just a wrapper?" No.

Most platforms wrap existing API calls and call it an 'agent'. We built the Lár Framework from scratch. It is a deterministic, graph-based execution engine designed for total state observability.
We didn't want a black box, so we built a glass one.

Production Grade Features

Multi-Provider

Switch between OpenAI, Gemini, and Anthropic with a single line of config.

Self-Correction

Built-in patterns for reflection and retry loops that catch mistakes.

Secure Logging

Firestore-first audit layers for enterprise-grade compliance.

Built with Lár

See the "Glass Box" philosophy in action with these open-source demos.

GlassBox RAG

A self-correcting RAG agent that plans, retrieves, drafts, and critiques its own answers. Watch it think in real-time.

Gemini 1.5 FAISS

Multi-Agent Support

A deterministic "assembly line" of agents. Triage -> Planner -> Specialist. No chaotic chat rooms, just pure orchestration.

Multi-Agent RouterNode

Lár: The PyTorch for Agents.