CRP Agent SDK¶

crp.Agent is the CRPv6 declarative agent surface. You declare tools + policy + model once and the protocol runs the positioned loop, executes tools, carries state forward, and emits governance metadata on every turn.

It is the fastest way to build agents that are provably grounded, auditable, and SLM-first.

One-line pitch¶

import crp

agent = crp.Agent(tools=[get_weather, convert_temp], profile="small-local")
result = agent.run("What's the weather in Sydney?")
print(result.answer)                 # natural-language answer
print(result.how_it_was_built)       # operation chain
print(result.crp.risk)               # LOW | MEDIUM | HIGH | CRITICAL
print(result.crp.grounded)           # True if sourced from tools/facts

Why `crp.Agent` over a raw tool loop¶

Raw tool loop	CRPv6 Agent SDK
You write prompt engineering, parser, retry, state, and safety glue.	Declare tools and policy; CRP builds the positioned frame.
Whole tool catalogue competes for context window.	Only the tools an operation needs are positioned per call.
Tool results are text; provenance is manual.	Every tool observation becomes a typed, grounded fact.
No built-in audit, safety, or continuation.	HMAC audit chain, safety surface, and continuation are automatic.
Works only on frontier models.	Same agent runs on 8B / 4B local models via LM Studio / Ollama.

Quick start¶

1. Install¶

pip install crprotocol

For local model support:

pip install crprotocol[full]

2. Declare an agent¶

from __future__ import annotations

import crp


def get_weather(city: str) -> str:
    """Return the current weather for a city."""
    return f"The weather in {city} is 22°C and sunny."


agent = crp.Agent(
    model="local:llama-3.1-8b",
    tools=[get_weather],
    system="You are a helpful weather assistant.",
    profile="small-local",
)

result = agent.run("What is the weather in Sydney?")
print(result.answer)

3. Inspect governance¶

print(result.crp.risk)            # LOW
print(result.crp.grounded)        # True
print(result.crp.chain_valid)       # True
print(result.how_it_was_built)    # retrieve
print(result.sources)             # tool observations with provenance

Execution profiles¶

The profile argument tells CRP how many capabilities to place in each Tool Positioning Frame (SPEC-049/050). Pick the one that matches your model.

Profile	Use case	Max tools per frame
`frontier`	GPT-4o, Claude, Gemini	7
`capable-local`	Qwen2.5-7B, Llama-3.1-8B	4
`small-local`	Qwen3-4B, Gemma-3-270M, Phi-4	2

Multi-turn state relay¶

Carry the Cognitive State Object (CSO) across turns so the agent remembers prior tool results without re-sending them in the prompt.

step1 = agent.run("What is the weather in London?")
step2 = agent.run("What was the temperature I just asked about?", prior_cso=step1.cso)

print(step2.answer)
print(step2.how_it_was_built)   # may include recall or retrieve

Local SLM proof¶

The CRPv6 Agent SDK has been proven live with LM Studio on:

meta-llama-3.1-8b-instruct
qwen3-4b

Example side-by-side demo:

python examples/crp_demos/live_llm_vs_crp.py

The script runs the same task with the same model and tools:

Raw LLM returns a JSON tool call and stops; the caller must parse and execute it.
CRPv6 Agent executes the tool, returns a natural-language answer, and emits governance metadata.

Full live proof (`live_crp_slm_proof.py`)¶

A comprehensive, readable demo that compares raw LLM and CRPv6 on four tasks:

Single tool
Tool chain
RAG retrieval
Long-form structured report

export CRP_LMSTUDIO_URL=http://localhost:1234/v1
export CRP_LMSTUDIO_MODEL=meta-llama-3.1-8b-instruct
python examples/crp_demos/live_crp_slm_proof.py

The script prints the full CRP governance block for every task, so you can see exactly what the protocol emits: risk, grounded, chain_valid, operations, and sources.

For a step-by-step video guide, see the Video Demo Guide.

Live test harness¶

python examples/crp_demos/live_agent_test_harness.py

Runs four agentic cases and reports pass/fail based on runtime governance metadata.

Agent templates¶

Ready-to-run templates are in examples/agents/:

Template	What it demonstrates	Provider
`slm_weather_agent.py`	Single + chained tools against a live local model	LM Studio / Ollama
`slm_rag_agent.py`	RAG retrieval against a live local model	LM Studio / Ollama
`weather_agent.py`	Same tools, mock provider (no API key needed)	Mock
`rag_agent.py`	RAG with mock provider	Mock
`gdpr_dsr_agent.py`	Sensitive workflow with checkpoint + audit	Mock
`report_agent.py`	Scan → summarise → report pipeline	Mock

The slm_*.py templates run against any OpenAI-compatible endpoint by setting CRP_LMSTUDIO_URL and CRP_LMSTUDIO_MODEL. The mock templates work out of the box for CI and documentation.

Additional reference templates are in examples/templates/:

customer_support_agent.py
code_review_agent.py
research_report_agent.py
data_analyst_agent.py
local_slm_agent.py

Reference¶

SPEC-049 - SLM Agent Execution Profile
SPEC-050 - Tool Capability Fabric & Operation Orchestration
SPEC-059 - Agent SDK
crp/agent_sdk/agent.py
crp/tools/capability_fabric.py
crp/stl/positioned.py