Give research agents a memory of what failed—and a way to test what works.
AgentQuant runs a research loop: propose an experiment, test it, inspect the result, and use that evidence to choose the next one. It tracks the whole search, including failed attempts, so you can compare research approaches at the same budget and replay their decisions.
Quantitative strategies are the first testbed. The broader question is simple: does the agent get better at choosing experiments, or does it just run more of them?
- Bounded experiments. Candidates share a backtest evaluator and cost model, with explicit search budgets.
- Memory with provenance. Trials, notes, and temporal graph evidence retain their sources and date cutoffs.
- Comparisons you can inspect. Compare search policies on the same episodes; export traces, episode bundles, and research memos.
- Tools around one research loop. Memory retrieval, hypothesis evidence, and candidate suggestions help the agent without owning evaluation or promotion.
Use it to study research-agent behavior, test memory and proposal policies, or build on an inspectable experiment loop.
Requires Python 3.10+.
git clone https://github.com/OnePunchMonk/AgentQuant.git
cd AgentQuant
python -m venv .venv
source .venv/bin/activate
python -m pip install -e .
python run_app.pyThe demo runs three search iterations on synthetic prices and writes results/demo_run.json. It shows the loop and its trace; it is not a trading-performance benchmark.
For the dashboard, install .[ui] and run python run_app.py --app. For LLM-guided proposals and web research, install .[llm] and configure provider credentials.
Walk through the research loop · Architecture · Memory
Run a comparison locally:
python scripts/fair_search_benchmark.py --episodes 3 --seeds 7 11 19 \
--output results/fair_search_benchmark.jsonThis defaults to synthetic fixtures. Offline prompt-mutation arms do not call an LLM, so matching scores are not evidence of learned improvement. See experimental details for interpretation.
The bounded search loop, memory service, replay artifacts, and advisory tools are implemented. Validated autonomous self-improvement remains a research goal: live policy comparisons, search-aware Sharpe inference, full promotion-gate integration, and prospective evaluation are still open.
For measured runtime improvements, see optimizations, including the graph-memory retrieval before-and-after diagram.
