lambdagent

lambdagentpaas — Lambda Calculus Agent Platform (monorepo)

lambdagentpaas is a full-stack platform for building, deploying, and serving AI agents defined in a Lambda-calculus DSL. The monorepo contains the language, the backend, the web UI, and ~10 example agents shipped to production.

📖 New here? Start with docs/INTRODUCTION.md for a project-level overview, then read this README for the DSL deep-dive.

v1.0 scope and deliberate non-goals

This first release prioritizes predictable, single-deployment usage over breadth. The audit (docs/AUDIT_2026-06-05.md) led to a series of subtractions before the first tag:

• Deleted: the Feishu bot integration (agentpaas/api/v1/feishu.py + agentpaas/services/feishu.py) — it was an unfinished dev-stage surface that handed process-wide bot-rebind capability to anyone reaching port 8000. Will return in v2 with proper signing.
• Deleted: agentexample/agent67/tools/shell_executor.py — LLM-controlled shell with a substring blocklist (RCE-via-prompt-injection). Was never wired into agent67's actual tool registry; the file is gone now too.
• Hardened: POST /api/v1/setup/bootstrap only accepts loopback callers. Production deploys behind a reverse proxy should use the agentpaas CLI to create the initial tenant.
• Single-tenant assumption: while the data model carries tenant_id columns and the audit found the scoping mostly correct, v1.0 is intended for single-tenant deployments. Multi-tenant requires the full RBAC + integration-test pass deferred to v2.
• Sample agents are samples: agentexample/qaagent67lambda ships with /home/67/knowledge/finance paths — treat as a reference implementation requiring path customization, not a turn-key template.

Things v1.0 is good for: running the lambdagent DSL kernel, executing YAML-defined agents through the PaaS API, building a knowledge base, using the web UI for chat/run/inspection, deploying via Docker Compose or the native scripts.

Things v1.0 is not yet good for: untrusted multi-tenant SaaS, knowledge-dir-as-attack-surface (no pickle for shared dirs), Feishu integration, or anything that needs the where = [".."] empty-wheel pre-fix tooling.

Repository layout

Directory	What it is	Run / install
lambdagent/	The core DSL — Lambda-calculus agent language, runtime, MCP/A2A/RAG/sandbox. Published as a standalone package (pip install lambdagent).	pip install -e lambdagent/
agentpaas/	PaaS layer — REST API, CLI, agent registry, run workspace, knowledge-base management. Wraps lambdagent for multi-tenant serving.	python3 -m agentpaas serve
webui/	React + Vite frontend — agent editor, chat, run inspector, knowledge-base UI.	cd webui && npm run dev
agentexample/	Concrete agents in production: physics67 (research pipeline), qaagent67* (RAG QA), travelagent67, pptagent67, research67, agent67, etc.	See each subdir's README
lambdagent_guard/	Static-analysis & runtime safety layer for compiled Lambda terms.	imported by agentpaas
deploy/	Native install + start/stop scripts (Bash, PowerShell, batch).	bash deploy/install-native.sh && bash deploy/start.sh
docker-compose.yml (repo root)	One-click Docker Compose deployment.	docker compose up -d
demo/	End-to-end demos (notebooks, scripts) for talks and onboarding.	—
docs/	Design docs, architecture, audits, comparison studies, INTRODUCTION.md.	—
tests/	189+ pytest suite across DSL, compiler, runtime, cost vectors, algebraic laws, AgentPack e2e.	pytest tests/

OntoRefactor data-governance platform

The monorepo now includes an end-to-end governance plane based on UML/MOF M3—M0 metamodeling. Business, software, and governance are composable M2 Profiles; important graph edges are reified assertions with evidence, confidence, validity, generator, and human-review state.

• Semantic source: ontology/
• LambdAgent pack: agentexample/agentpacks/governance.ontology-steward/
• REST API: /api/v1/governance/* (15 paths, tenant scoped)
• Web UI: open 数据治理 in the sidebar, then load the built-in tenant-isolation sample
• Inputs: MySQL DDL, OpenAPI YAML/JSON, or JSON asset inventory
• Workflow: deterministic extraction → parallel LambdAgent analysis → M0 facts/M1 candidates → validation → evidence/review → impact graph

After starting AgentPaaS and the Web UI, the fastest demo is:

1. Open 数据治理 and click 加载租户隔离样例.
2. Inspect the 48-node M3—M0 reference slice and its cross-Profile relations.
3. Paste DDL into 智能体分析 to create evidence-backed facts and candidates.
4. Review inferred assertions, run the governance gates, and inspect impact paths.

Live deployments

Service	URL	Notes
maritime-qa	http://qa.lambdagent.cn:8080	海事法规 RAG
投标-qa	http://qa.lambdagent.cn:8084	投标文件 RAG
clean-qa	http://qa.lambdagent.cn:8083	49 docs / 1527 chunks, BM25 + Graph + Wiki

---

About the DSL (rest of this document)

The remainder of this README documents the lambdagent/ subpackage — the Lambda-calculus DSL itself. For platform usage (REST API, web UI, deploying your own agent), see docs/INTRODUCTION.md and AGENT_GUIDE.md.

A Python DSL that models AI agents as Lambda calculus terms. Every agent is a function, every composition is function composition, every loop is a Y combinator. This is not a metaphor — it is a formal correspondence backed by a Church encoding verification.

Key insight: An (M, D) pair — a language model M plus a dataset/prompt D — is isomorphic to a Lambda term. LLM + Dataset = Lambda term.

Stats: ~13,500 lines of Python | 197 exported symbols | 4 patents filed

Highlights

• Unified LLM Provider System -- 4 providers cover all major backends: Claude Code (via Max Plan), Anthropic API, OpenAI-compatible (DashScope / DeepSeek / Moonshot / Zhipu), and Ollama (local). Switch with one line of YAML.
• Zero-Hallucination Conversations -- ConversationLam maintains full session persistence so the model never loses context mid-conversation, eliminating the hallucination caused by truncated history.
• Claude Code Max Plan -- No API Key -- Set provider: claude-code and use your Claude Code subscription directly. No API key, no billing dashboard, no environment variables.
• YAML-Driven Configuration -- Define agents, tools, memory, and provider settings in a single YAML file. from_config() compiles it into a typed Lambda term tree.

Quick Start (Simplest Path)

# agent-config.yml
model:
  provider: claude-code
  name: sonnet

# Interactive chat
python3 -m agentpaas chat lambda

# Or run the example agent directly
python3 agentexample/agent67/run.py --claude

No API key required -- uses your Claude Code Max Plan subscription.

Provider Comparison

Provider	Config value	API Key needed	Notes
Claude Code (Max Plan)	claude-code	No	Uses local Claude Code subscription
Anthropic API	anthropic	Yes (ANTHROPIC_API_KEY)	Direct Claude API access
OpenAI-compatible	dashscope / openai / deepseek / zhipu / moonshot	Yes	Any OpenAI-compatible endpoint
Ollama (local)	ollama	No	Runs models locally, requires Ollama installed

Features

Core Constructs (Lambda Calculus)

lambdagent maps 11 core agent constructs to Lambda calculus, giving each one a precise denotational semantics:

#	Lambda Calculus	DSL Construct	Description
1	Lambda abstraction λx.body	Lam(name, prompt)	An LLM with a system prompt
2	Application (f x)	agent(input)	Run agent on input
3	Composition λx.g(f(x))	f >> g	Pipeline / chain
4	Church conditional IF	If(cond, then_, else_)	Conditional branching
5	Y combinator	Loop(body, condition)	Iterative reasoning (ReAct)
6	Church pair PAIR	Pair(f, g)	Run two agents, return both
7	Projections FST / SND	Fst() / Snd()	Extract from pair
8	Primitive / Oracle	Tool(name, fn)	External tool call
9	Generalized CASE	Route(classifier, routes)	Multi-way dispatch
10	Dependent type {x:T\|P(x)}	Guard(agent, validator)	Output validation
11	Environment extension	Memory(agent, store)	Persistent context

Plus Par(f, g) as syntactic sugar for parallel execution (via | operator).

Multi-Agent Constructs (pi-calculus extension)

#	Process Calculus	DSL Construct	Description
12	Channel c!(v) / c?(x)	Channel + Send + Receive	Inter-agent communication
13	Shared environment	SharedMemory	Thread-safe shared state
14	Y_n(Loop + Route)	GroupChat	Multi-agent group discussion
15	Dynamic CASE	Handoff	Runtime dynamic delegation
16	Concurrent beta-reduction	AsyncPar	Thread-pool true parallelism

Skill System

Construct	Description
Skill	Named, reusable Lambda term with metadata
SkillPack	Collection of related skills (like a package)
SkillRegistry	Global singleton registry with search/discover
SkillAgent	Auto-discovers and executes best skill
@skill	Decorator to create and register skills

Protocol Integrations

Module	Protocol	Description
MCPServer/Tool	MCP 2025-11-25	Connect to MCP servers (HTTP + stdio)
A2AServer/Client	Google A2A v0.3	Publish/discover/call remote agents
RAGTool	--	TF-IDF or ChromaDB retrieval-augmented gen
AgenticRAG	--	Agent decides when to retrieve
Checkpoint	--	Serialize/restore execution state to JSON

Sandbox (Process Isolation)

Construct	Description
SandboxedTool	Tool that runs in isolated subprocess with resource limits
SandboxPolicy	Security policy (timeout, memory, network, etc.)
SecureExecutor	Auto-wraps all Tools in a term tree with sandbox
ResourceLimiter	Applies CPU/memory/fd limits via POSIX resource
@sandboxed	One-line decorator for sandboxed tool creation

Architecture

YAML Config ──→ from_config() ──→ Lambda Term ──→ Runtime ──→ Result
                 (compiler)        (Term tree)    (beta-reduction)

                               ┌── Multi-Agent: Channel, GroupChat, Handoff, AsyncPar
                               ├── Skills: Skill, SkillPack, SkillRegistry, SkillAgent
Lambda Term ──→ Runtime ──→    ├── MCP: MCPServer, MCPTool (HTTP + stdio)
                               ├── A2A: AgentCard, A2AServer, A2AClient
                               ├── RAG: RAGTool, AgenticRAG, SimpleVectorStore
                               ├── Checkpoint: save/load execution state
                               └── Sandbox: SandboxedTool, SecureExecutor, ResourceLimiter

The system has five layers:

1. Core DSL (lambdagent.core, lambdagent.primitives, lambdagent.extensions) — the 11 core constructs as Python classes, each a subclass of Term.
2. Multi-Agent (lambdagent.multiagent) — 5 pi-calculus constructs for inter-agent communication, group chat, dynamic delegation, and true parallelism.
3. Skill + Protocol (lambdagent.skills, lambdagent.mcp_client, lambdagent.a2a, lambdagent.rag, lambdagent.checkpoint) — reusable skills, MCP/A2A protocol integration, retrieval-augmented generation, and state persistence.
4. Compiler (lambdagent.fromconfig) — parses YAML configs into Lambda term trees. Includes schema validation and lint.
5. Runtime (lambdagent.agentruntime) — executes Lambda terms via beta-reduction, handling LLM calls, MCP tool invocation, memory, and tracing. 支持双执行引擎切换 (recursive/cek/adaptive), 通过 runtime.engine 配置。CEK 引擎提供逐步成本监控、暂停/恢复和循环检测能力。

Installation

pip install lambdagent

For LLM provider support:

pip install lambdagent[anthropic]   # Anthropic Claude
pip install lambdagent[openai]      # OpenAI-compatible (Dashscope, etc.)

Quick Start

Python DSL

from lambdagent import Lam, Compose, Loop, Tool, Memory

# A single agent is a Lambda abstraction
writer = Lam("writer", "You are a technical writer. Write clear documentation.")

# Composition is function composition
reviewer = Lam("reviewer", "Review the text for clarity and correctness.")
pipeline = writer >> reviewer  # = lambda x. reviewer(writer(x))

# ReAct agent is a Y combinator
# `condition` takes (result, step_index) and returns bool;
# `max_steps` is the unfolding bound (default 10).
searcher = Tool("search", lambda q: web_search(q))
researcher = Loop(
    body=Lam("think", "Analyze the question. Use search if needed."),
    condition=lambda result, step: "DONE" in result,
    max_steps=10,
)

# Add memory for persistent context.
# `store` is an optional dict of {key: value} pre-loaded into the agent's
# environment; further entries are added via `agent.remember(k, v)`.
agent = Memory(researcher, store={"recent_topic": "LLM agents"})

# Execute = beta-reduction
result = agent("What are the latest advances in LLM agents?")

YAML Configuration

agentId: research-assistant
name: ResearchAssistant
type: react
systemPrompt: |
  You are a research assistant. Search for information,
  analyze it, and produce a comprehensive report.
model:
  provider: dashscope
  name: qwen3-max-2026-01-23
  temperature: 0.7
  maxTokens: 4096
react:
  maxSteps: 15
runtime:
  engine: recursive        # recursive | cek | adaptive
memory:
  enabled: true
  strategy: local
  size: 20
mcp:
  onlineTool:
    example-mcp-server:
      - everything_get_sum
  localTools:
    - terminate

from lambdagent import from_config

agent = from_config("agent-config.yml")
result = agent("Summarize recent progress in AI safety research")

Multi-Agent Group Chat

from lambdagent import Lam, GroupChat, SharedMemory

researcher = Lam("researcher", "You are a researcher. Find evidence.")
critic = Lam("critic", "You are a critic. Challenge weak arguments.")
synthesizer = Lam("synthesizer", "You synthesize the discussion into conclusions.")

chat = GroupChat(
    agents=[researcher, critic, synthesizer],
    max_rounds=6,
    scheduler="round_robin",
)
result = chat("Should we invest in quantum computing?")

MCP Tool Integration

from lambdagent import mcp_tools, mcp_tool

# Get all tools from an MCP server
tools = mcp_tools("http://localhost:3000/mcp")

# Or get a single tool
search = mcp_tool("http://localhost:3000/mcp", "search")
result = search({"query": "AI agents"})

RAG (Retrieval-Augmented Generation)

from lambdagent import create_rag, AgenticRAG, Lam

rag = create_rag(["Python is a programming language.", "Lambda calculus is..."])
agent = Lam("qa", "Answer questions using the provided context.")
agentic = AgenticRAG(agent, rag, decider=lambda x: "?" in x)
result = agentic("What is Lambda calculus?")

Checkpoint (Save/Resume)

from lambdagent import Context, save_context, load_context

ctx = Context()
agent("long running task", ctx)
save_context(ctx, "checkpoint.json", last_input="long running task")

# Later, resume:
ctx = load_context("checkpoint.json")
agent("continue from here", ctx)

Sandbox (Process Isolation)

from lambdagent import SandboxedTool, SandboxPolicy, SecureExecutor, sandboxed

# One-line decorator
@sandboxed(timeout=10, memory_mb=128)
def risky_calc(x):
    return eval(x)

# Or explicit construction with policy presets
tool = SandboxedTool("calc", lambda x: eval(x), policy=SandboxPolicy.strict())

# Secure an entire term tree — wraps all Tools in sandboxes
secure_tree = SecureExecutor(policy=SandboxPolicy.default()).sandbox_all_tools(agent)

One-Sentence Agent Builder

python experiments/nl2agent.py "Build a research assistant that can search the web and write reports, up to 20 steps, with memory"

Or programmatically:

from experiments.nl2agent import one_sentence_to_agent

one_sentence_to_agent(
    "Build a code review pipeline: check security, then style, then performance",
    user_input="Review this Python function: def login(u, p): ...",
)

CLI Usage

# Compile YAML to Lambda term (no execution)
lambdagent compile agent-config.yml

# Compile and execute
lambdagent run agent-config.yml "Write a quicksort in Python"

# Interactive REPL
lambdagent repl agent-config.yml

# Static analysis (lint)
lambdagent lint agent-config.yml

# Export pure Lambda expression
lambdagent lambda agent-config.yml

# Unix pipes work (composition = pipe)
echo "Hello world" | lambdagent run agent-config.yml -

API Reference

Core Terms (11 constructs)

• Term — Abstract base class. All constructs are Terms.
• Context — Evaluation environment with beta-reduction tracing.
• Lam(name, prompt) — Lambda abstraction. Wraps an LLM call.
• Compose(f, g) / f >> g — Function composition.
• If(cond, then_, else_) — Church conditional.
• Loop(body, condition) — Y combinator with termination.
• Pair(f, g) / Par(f, g) — Church pair / parallel execution.
• Fst() / Snd() — Pair projections.
• Tool(name, fn) — External oracle (MCP tool, shell command, etc.).
• Route(classifier, routes) — Generalized Church boolean (CASE).
• Guard(agent, validator) — Dependent type / output validation.
• Memory(agent, store) — Environment extension with persistent state.

Multi-Agent (5 constructs)

• Channel(name, capacity) — pi-calculus channel for inter-agent communication.
• Send(agent, channel) — Send agent output to channel.
• Receive(channel, handler) — Receive from channel, optionally process with handler.
• SharedMemory(store, append_only) — Thread-safe shared state across agents.
• GroupChat(agents, scheduler, max_rounds) — Multi-agent group discussion (Y + Route).
• Handoff(selector, registry, fallback) — Runtime dynamic delegation.
• AsyncPar(*agents) — True parallel execution via thread pool.

Skill System

• Skill(name, term, description, signature, tags) — Named, reusable Lambda term with metadata.
• SkillPack(name) — Collection of related skills.
• SkillRegistry() — Global singleton registry (search, discover, build_route).
• SkillAgent(classifier, registry) — Auto-discovers and executes best skill.
• @skill(name, description, tags) — Decorator to create and auto-register skills.

MCP Client

• MCPServer.http(url) / MCPServer.stdio(command) — Connect to MCP servers.
• MCPTool(server, tool_name) — MCP tool wrapped as lambdagent Term.
• mcp_tools(url) — One-liner to get all tools from a server.
• mcp_tool(url, name) — One-liner to get a single tool.

A2A Protocol

• AgentCard — A2A Agent capability description (JSON).
• A2AServer(agent, port) — Publish agent as A2A HTTP service.
• A2AClient(url) — Call remote A2A agent as local Term.
• skill_to_agent_card(skill) — Convert Skill to AgentCard.

RAG

• RAGTool(store, top_k) — Retrieval tool (TF-IDF or ChromaDB).
• AgenticRAG(agent, rag, decider) — Agent decides when to retrieve.
• SimpleVectorStore() — Zero-dependency TF-IDF vector store.
• create_rag(documents, top_k) — One-liner to create RAG tool.

Checkpoint

• Checkpoint(context, shared_data) — Execution state snapshot.
• CheckpointManager(directory) — Manage multiple checkpoints with auto-cleanup.
• save_context(ctx, path) / load_context(path) — Save/restore Context.

Sandbox

• SandboxedTool(name, fn, policy) — Tool running in isolated subprocess with resource limits.
• SandboxPolicy(timeout, memory_mb, network, ...) — Security policy. Presets: .strict(), .default(), .permissive().
• SecureExecutor(policy) — Auto-wraps all Tools in a term tree via sandbox_all_tools().
• ResourceLimiter — Applies POSIX RLIMIT_CPU, RLIMIT_AS, RLIMIT_NOFILE, RLIMIT_NPROC.
• @sandboxed(timeout, memory_mb, ...) — Decorator for one-line sandboxed tool creation.
• Exceptions: SandboxViolation, TimeoutViolation, MemoryViolation, OutputViolation.

Compiler

• from_config(path_or_dict) — YAML to Lambda term.
• lint_config(path_or_dict) — Static analysis.
• to_lambda_expr(path_or_dict) — Export formal Lambda expression.
• describe_config(path_or_dict) — Human-readable structure description.

Runtime

• Runtime — Executes Lambda terms with LLM backends.
• ReActEngine — Implements the Y combinator loop for ReAct agents.
• MCPClient — Model Context Protocol tool invocation.
• TraceStore — Records beta-reduction traces for debugging.

Theory

lambdagent is grounded in a formal correspondence between agent constructs and Lambda calculus:

• Church encoding verification: All 11 constructs are shown to be encodable as pure Lambda terms. The If construct uses Church booleans, Pair/Fst/Snd use Church pairs, and Route uses generalized Church numerals.
• S+K completeness: The DSL is computationally complete — Lam and application together can encode S and K combinators, from which all computable functions follow.
• Y combinator semantics: Loop is not ad-hoc iteration; it is the Y combinator Y = λf.(λx.f(x x))(λx.f(x x)) with a bounded unfolding (max_iterations) to ensure termination.

For the full formal treatment, see docs/INTRODUCTION.md and the design docs under docs/ (Church-encoding proofs, cost-vector semantics, algebraic laws).

License

Business Source License 1.1 — Copyright (c) 2025 kenny67nju

Non-production use (development, testing, personal projects, academic research) is always permitted. Production use is free for organizations with ≤10 individual users or employees. On 2031-04-05 (the Change Date), the license automatically converts to Apache License 2.0.

For production use beyond the 10-user limit before the Change Date, contact the licensor for a commercial license.

Citation

If you use lambdagent in academic work, please cite:

@software{lambdagentpaas2025,
  title     = {lambdagentpaas: A Lambda Calculus Agent Platform},
  author    = {kenny67nju},
  year      = {2025},
  url       = {https://github.com/kenny67nju/lambdagentpaas},
}

To cite the DSL specifically (the lambdagent/ subpackage, also published standalone on PyPI):

@software{lambdagent2025,
  title     = {lambdagent: A Lambda Calculus Agent DSL},
  author    = {kenny67nju},
  year      = {2025},
  url       = {https://github.com/kenny67nju/lambdagent},
}