AgentQuant AI Agent

AgentQuant

A reproducible, failure-aware harness for research agents. AgentQuant lets an agent propose experiments, test them under a fixed budget, inspect failures, and compare changes to its own research policy. Quantitative strategies are the first testbed.

The goal is to answer a specific question: does evidence from earlier experiments help the agent choose better next experiments on unseen episodes, at the same cost? A changed prompt or a higher in-sample score alone is not self-improvement.

Try it locally

python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[dev]"
python run_app.py

This zero-key demo uses deterministic synthetic prices and a grid/random fallback. It checks that the loop runs and produces a report in results/; it does not measure trading performance. For the optional dashboard, install .[ui] and run python run_app.py --app. LLM-guided proposals and web research require .[llm] and provider credentials.

How it works

Edit the diagram in Excalidraw.

• Search: a bounded loop proposes strategy candidates, applies one backtest evaluator with costs, and records failed as well as successful attempts.
• Memory: SQLite stores trials and retrieves evidence available by the market-time cutoff. A temporal evidence graph is an experimental, rebuildable projection for relationship-based discovery.
• Policy improvement: an outer loop can propose changes to the research process. A candidate must clear protected and statistical evaluation before promotion. The standalone search-aware gate is a foundation; validated Sharpe inference and full application integration remain open work.
• Replay: manifests, episode bundles, traces, and research memos make decisions inspectable. See the architecture and implementation handoff for current wiring and remaining work.

The assistant tool contract explains how idea, hypothesis, and memory sidecars expose bounded, auditable calls.

Reproduce development comparisons

python scripts/fair_search_benchmark.py --episodes 3 --seeds 7 11 19 \
  --output results/fair_search_benchmark.json
python scripts/bounded_self_improvement.py --episodes 6 --seeds 7 11 19 \
  --n-mutations 3 --output results/bounded_self_improvement.json
python scripts/export_research_memo.py --episodes 6 --seeds 7 11 19 \
  --n-mutations 3 --output results/research_memo

These default to synthetic fixtures. In offline fallback mode, prompt mutation does not reach an LLM; identical mutation-arm results cannot support an evidence-guided improvement claim. The live comparison in #31, statistical work in #48, and prospective evaluation are still required.

Explore the project

• Architecture and agent implementation handoff
• Worked research loop and memory design
• Experimental details and research memo design
• Assistant tool demo, communication experiment, and benchmark evolution
• Harnesskit trace integration and Peek leakage auditor
• Contributing, roadmap, and open issues

Backtests are research measurements, not predictions of future returns. The repository does not claim a profitable strategy or validated autonomous self-improvement. License: MIT.