End the
Black Box Tax.
A commitment to transparent, deterministic, and auditable AI — grounded in six published papers and enforced at the execution layer.
You're a developer. Your agent fails in production.
What do you get? A 100-line stack trace. No state. No inputs. No visibility.
This is the Black Box Tax.
The silent penalty paid when systems hide the truth from the people responsible for them.
The Cost of Obscurity
| Feature | "Black Box" Frameworks | Lár — The Glass Box |
|---|---|---|
| Debugging | Guesswork. 100-line stack traces from inside a "magic" executor. | Precise. See the exact node, state, and error that caused the failure. |
| Auditability | Paid add-on. Requires external tools to trace execution. | Built-in. The Flight Log is the core output of the engine. |
| Control | Chaotic. Agents "chat" to pass data. Order is unpredictable. | Deterministic. You define the assembly line. Data flow is explicit. |
| Security | Cloud-locked. Your data leaves the perimeter to hit their servers. | Zero telemetry. The only egress is to your LLM provider. |
| EU AI Act | Compliance is a wrapper. "Trust us" is the audit trail. | 13 structural primitives. Art. 12, 13, 14, 15(4), GDPR — enforced in code. |
The Flight Log
Lár produces a Flight Log for every run. Not a debug tool — a legal record of your AI's decisions. Every entry is HMAC-SHA256 signed. One byte changed post-run: verification fails.
{
"steps": [
{
"step": 0,
"node": "LLMNode",
"state_before": {
"task": "What is the Lár Framework used for?"
},
"state_diff": {
"added": {
"category": "GENERAL"
},
"removed": {},
"modified": {}
},
"run_metadata": {
"prompt_tokens": 45,
"output_tokens": 449,
"total_tokens": 494,
"model": "gemini/gemini-2.5-pro"
},
"outcome": "success"
},
{
"step": 1,
"node": "LLMNode",
"state_before": {
"task": "What is the Lár Framework used for?",
"category": "GENERAL",
"__last_run_metadata": null
},
"state_diff": {
"added": {
"search_query": "Lár Framework use cases"
},
"removed": {
"__last_run_metadata": null
},
"modified": {}
},
"run_metadata": {
"prompt_tokens": 83,
"output_tokens": 1717,
"total_tokens": 1800,
"model": "gemini/gemini-2.5-pro"
},
"outcome": "success"
},
{
"step": 2,
"node": "ToolNode",
"state_before": {
"task": "What is the Lár Framework used for?",
"category": "GENERAL",
"__last_run_metadata": null,
"search_query": "Lár Framework use cases"
},
"state_diff": {
"added": {
"retrieved_context": "=== Lár Engine Knowledge Base ===\n\nProduct: lar-engine (The Open-Source Framework)\n\nTopic: General Questions\n\nQuestion: What is Lár?\n\nAnswer: Lár (Irish for \"core\") is our open-source, \"glass box\" agentic framework. It is the \"PyTorch for Agents.\" It's a simple, \"dumb\" engine that lets you build, run, and audit complex AI agents one step at a time.\n\nQuestion: What is the \"Glass Box\" philosophy?\n\nAnswer: \"Glass Box\" means 100% auditability. Our lar engine's core output is a step-by-step log of every state change. Unlike \"black box\" frameworks that hide their logic, lar lets you see exactly why your agent failed, which node was responsible, and what data it was processing.\n\nQuestion: How do I get support?\n\nAnswer: Please check our GitHub repositories.\n\nTopic: Licensing\n\nQuestion: How much does lar cost?\n\nAnswer: The lar-engine is, and always will be, 100% free and open-source under an Apache 2.0 license."
},
"removed": {
"__last_run_metadata": null
},
"modified": {}
},
"run_metadata": null,
"outcome": "success"
},
{
"step": 3,
"node": "RouterNode",
"state_before": {
"task": "What is the Lár Framework used for?",
"category": "GENERAL",
"__last_run_metadata": null,
"search_query": "Lár Framework use cases",
"retrieved_context": "..."
},
"state_diff": {
"added": {},
"removed": {
"__last_run_metadata": null
},
"modified": {}
},
"run_metadata": null,
"outcome": "success"
},
{
"step": 4,
"node": "LLMNode",
"state_before": {
"task": "What is the Lár Framework used for?",
"category": "GENERAL",
"__last_run_metadata": null,
"search_query": "Lár Framework use cases",
"retrieved_context": "..."
},
"state_diff": {
"added": {
"agent_answer": "Lár (Irish for \"core\") is our open-source, \"glass box\" agentic framework. It is the \"PyTorch for Agents.\" It's a simple, \"dumb\" engine that lets you build, run, and audit complex AI agents one step at a time."
},
"removed": {
"__last_run_metadata": null
},
"modified": {}
},
"run_metadata": {
"prompt_tokens": 689,
"output_tokens": 761,
"total_tokens": 1450,
"model": "gemini/gemini-2.5-pro"
},
"outcome": "success"
},
{
"step": 5,
"node": "AddValueNode",
"state_before": {
"task": "What is the Lár Framework used for?",
"category": "GENERAL",
"__last_run_metadata": null,
"search_query": "Lár Framework use cases",
"retrieved_context": "...",
"agent_answer": "Lár (Irish for \"core\") is our open-source, \"glass box\" agentic framework. It is the \"PyTorch for Agents.\" It's a simple, \"dumb\" engine that lets you build, run, and audit complex AI agents one step at a time."
},
"state_diff": {
"added": {
"final_response": "Lár (Irish for \"core\") is our open-source, \"glass box\" agentic framework. It is the \"PyTorch for Agents.\" It's a simple, \"dumb\" engine that lets you build, run, and audit complex AI agents one step at a time."
},
"removed": {
"__last_run_metadata": null
},
"modified": {}
},
"run_metadata": null,
"outcome": "success"
}
],
"summary": {
"total_steps": 6,
"total_prompt_tokens": 817,
"total_completion_tokens": 2927,
"total_tokens": 3744
}
}Every decision. Every state change. Signed, timestamped, verifiable.
The compliance guarantees are published.
Not configured. Not asserted. Proven.
Six papers in the Lár Research Series establish the formal properties that every Lár deployment inherits: 33 named invariants, annotation-free continual learning, no catastrophic forgetting by design. Permanent Zenodo DOIs — not blog posts.
Is this just a wrapper? No. Most platforms wrap APIs and call it an "agent". Lár is a deterministic execution graph built from scratch. It uses LiteLLM for universal model support (100+ providers), but the execution graph is pure, debuggable Python designed for total state observability. The research series formally characterises every architectural property it claims.
Verify the code on GitHub →The Core Primitives
We rejected complexity. No magic. Just Python.
Composable building blocks. Combine them to build any compliant agent.
Time Travel Debugging
Lár logs State Diffs, not full state copies. The GraphExecutor yields a lightweight step_log that shows exactly what changed — infinite scalability, perfect replayability.
Audit Log Internals →
Show Your Agents Are Auditable
If you build an agent using the Lár Engine, you're building a verifiable system. Display the badge to signal that your agent is built for production reliability and auditability.
Contribute to the Core
Lár is open source. The Glass Box philosophy is built with the community. If you want to help shape the future of agentic auditability, join us.
Read Contributing GuidelinesReady to break the Black Box?
Pure Python. No magic. No license fees for auditability.