MixPilot

An agentic harness for marketing measurement.

Everyone is building general-purpose agents. MixPilot is the opposite bet: a small, real agent runtime whose entire action space is the marketing-science toolkit — adstock, saturation, multi-touch attribution, budget optimisation, and a marketing-aware data-quality audit — wired as typed tools with structured results, an approval gate, and an eval suite that scores the agent's method-selection judgment.

The thesis: in a domain agent, the tools are the product. A model can sound confident about marketing data; the hard part is knowing which method the data can actually support — and refusing the ones it can't. That judgment is what MixPilot encodes and tests.

pip install mixpilot            # tools only
pip install "mixpilot[groq]"    # + the agent loop on Groq (open models)
pip install "mixpilot[agent]"   # + the agent loop on Anthropic

The agent loop is provider-agnostic — the LLM client is injected, so the same loop runs on Groq Llama or Claude by swapping one line.

What's inside

The judgment, made concrete

Same data, two attribution models:

from mixpilot.tools.attribution import run_attribution_model

paths  = [{"path": ["Social", "Search"], "converted": True}] * 40
paths += [{"path": ["Social"],           "converted": False}] * 20

run_attribution_model(paths, "last_touch").artifacts["share"]
# {'Search': 1.0, 'Social': 0.0}   <- the assist is thrown away
run_attribution_model(paths, "markov").artifacts["share"]
# {'Search': 0.5, 'Social': 0.5}   <- both channels are necessary

Last-touch hands Social nothing. Markov's removal effect sees that every conversion needed Social first, and splits the credit. That gap — assist value — is the entire reason multi-touch attribution exists.

Running the agent

from mixpilot import Agent
from mixpilot.agent.groq_client import GroqClient

agent = Agent(GroqClient(model="llama-3.3-70b-versatile"))  # needs GROQ_API_KEY
result = agent.run(
    "Here are 12 weeks of spend and sales for one channel ... "
    "what's the saturation point and should we spend more?"
)
print(result.final_text)

Swap GroqClient for AnthropicClient (needs ANTHROPIC_API_KEY) to run the same loop on Claude. The client is injected, so production uses a real model and the evals use a scripted client — no network needed to test the harness.

Evals: scoring method selection

python evals/run_evals.py          # offline, deterministic
python evals/run_evals.py --live   # against a real model

Each case pairs a data shape with the methods it does and does not support. A case passes only if the agent uses every required tool and avoids every forbidden one — e.g. it must not run Markov attribution on single-touch data, and must audit before claiming per-channel effects on collinear spend.

Tests

python -m pytest -q     # 14 passing — domain math, harness contracts, provider translation

The tests never call a model. They prove the adstock conserves mass, the Hill fit recovers known parameters, the Markov removal effect credits assists, the optimiser moves budget toward unsaturated channels, the audit catches multicollinearity, and the approval gate blocks mutating actions. Harness reliability that has nothing to do with model intelligence.

Scope

This is deliberately small. No context compaction, no MCP, no subagents — those are solved problems in general harnesses. The point here is the opposite: how far you get when the action space is a real domain and the tools enforce the judgment.

MIT licensed.