You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
MartinLoop is an open-source control layer for AI coding agents.
AI agents are starting to write code, run tools, edit files, and make decisions with limited human supervision. Today, they can loop, overspend, make unverified changes, or claim success without clear evidence. MartinLoop adds practical safety rails: budget caps, verification gates, retry limits, rollback, and auditable run records.
The goal is to make autonomous coding agents safer, more bounded, and easier to inspect before they become default software workers.
Improve the MartinLoop open-source runtime.
Publish a clear agent failure taxonomy.
Build reproducible benchmark examples showing common agent failure modes.
Document how budget caps, verification, rollback, and audit trails reduce risk.
Make the tool easier for developers and AI safety researchers to test.
I will achieve this by tightening the runtime, running evaluation examples, improving documentation, and publishing concrete case studies of agent runs with and without controls.
The funding will support a focused 3–4 month push to turn MartinLoop into stronger open-source AI safety infrastructure for autonomous coding agents.
With the $25,000 minimum, I would fund:
founder build time for core runtime improvements
basic API/model evaluation runs
improved documentation
an initial agent failure taxonomy
a small benchmark showing failures with and without controls
With the $60,000 full goal, I would fund:
$30,000 founder stipend for focused build/research time
$10,000 API/model credits for benchmark and evaluation runs
$5,000 cloud/dev infrastructure, CI, logging, and hosted demos
$5,000 engineering support for benchmark harness and runtime improvements
$5,000 security/reliability review of rollback, logging, and audit trails
$3,000 documentation, diagrams, and onboarding materials, benchmark development
$2,000 contributor bounties, user testing, and admin tools
This would produce a better runtime, public benchmarks, clearer documentation, and concrete examples showing how budget caps, verification, rollback, and audit trails reduce agent failure modes.
Vakeesan Mahalingam (myself) and Gobi Shantha.
I am a 5x founder, ex-VC, and CFA charterholder who loves to bring ideas to life. I built and published MartinLoop as an open-source project.
My background includes product strategy, technical execution, financial risk/control thinking, and public-facing technical communication. MartinLoop builds on my interest in practical AI safety: making agent systems bounded, verifiable, auditable, and recoverable.
Gobi is a full stack SWE with 10+ years of applied experience across multiple languages, react, nodejs, typescript, c++, rust, and loves back end work.
We've work on 3 companies together (2 AI and a VC backed Bitcoin infra company)
The main failure modes are:
Developers may not adopt another tool unless integration is very lightweight.
The benchmarks may not be convincing enough to show safety value.
The project may remain useful as a niche open-source tool but not become broader safety infrastructure.
If it fails, the likely outcome is still useful: public documentation, agent failure examples, and an open-source prototype others can learn from or build on. I will reduce this risk by keeping the scope focused, publishing concrete examples, and prioritizing developer usability.
No dedicated grant funding has been raised for MartinLoop in the last 12 months. The project has been self-funded through founder time and related startup work and frontier model API credits.