Independent work · Experimental agent system
Meet frAInk.
An autonomous AI agent built in the image of its creator. Real code. Real decisions. Real guardrails. Progressive autonomy earned through performance.
frAInk started as an experiment: what happens when you give an AI agent real money, real guardrails, and let it learn from outcomes? It grew into a system that writes and ships its own code, manages a team of specialist agents, and earns expanded capability based on validated results. Nothing is faked. Everything is logged.
Architecture
Three roles. None of them trusts the others.
PLANNER
Read-only research layer. Fetches, analyzes, synthesizes. Never decides. Never acts.
POLICY
The gatekeeper. Evaluates reports, enforces guardrails, approves or blocks. Never executes.
EXECUTOR
The action layer. Touches APIs, sends messages, runs builds. Only acts on Policy-approved instructions.
Guardrails are non-negotiable. Budget caps per cycle. Kill switch. Policy approval required for every action. Everything logged. Anomalies surfaced. Nothing buried.
Capabilities
What it does — and what it did.
Coding SDK
frAInk identifies tasks in its own codebase, proposes builds, writes the code, runs tests, opens PRs, and merges — all autonomously. Triggered from Telegram, executed by Sonnet, reviewed by the pipeline.
Multi-Agent Management
frAInk delegates to a team of specialist agents — code review, cost auditing, capability design, content evaluation, and brain curation. Each specialist has its own scope, budget ceiling, and quality metrics. Auto-demotion on regression.
Self-Build Loop
A weekly cycle where frAInk proposes new tools, validates them in a sandbox, and promotes them to production. It proposes behavioral rules from its own experience, shadows them against outcomes, and promotes rules that prove their value.
Content & Public Presence
Voice corpus extraction from real conversations. Lab publisher that auto-syncs experiment results to the public repo. Every public artifact is generated from real work — not manufactured engagement.
Autonomous Trading Retired
Built and ran paper and live trading on Alpaca with a multi-factor confluence screener, market regime detection, and position management. Every trade decision flowed through the full Planner → Policy → Executor pipeline. The trading lane was retired in July 2026; the screener and policy machinery remain.
Prediction Markets Retired
Historically: live Kalshi integration with RSA-PSS authentication, cross-referencing public forecasts and consensus against market pricing. Real money, real wins, real losses — all published in the lab repo. Retired with the trading lane in July 2026.
Evolution
How it got here
- Phase 1 — Foundation
3-agent pipeline, first live pipeline run, Planner/Policy/Executor architecture proven.
- Phase 2 — Capability Expansion
Web search, financial analysis tooling, article evaluator, paper trading, prediction-market integration, intelligence briefings.
- Autonomy Weekend
20-hour build: memory schema, rules engine, shadow mode, proposal scanner, rule proposer, tool proposer, self-build loop operational.
- Phase 3 — Multi-Agent Delegation
Specialist agents registered, task queue with priority dispatch, quality regression flow with auto-demotion, progressive trust expansion.
- Coding SDK — frAInk Ships Code
Autonomous task identification, code generation, test coverage, PR creation, and merge. First pilot build: $2.57, 25 turns, PR merged.
- Trading Lane Retired
Trading lane retired; build/coding lanes continue. Published experiment results stay in the lab repo as history.
The origin. Frank Kronstein → Frankenstein → frAInk. Public experiment files and build narratives live in github.com/letmebefraink/frAInk-lab — full hypotheses, guardrails, raw results, and build docs.