# Cashew Labs: self-improving agent systems

Canonical: https://mudpie.ai/companies/cashew-labs/
Breadcrumb: [Home](https://mudpie.ai/) / [Companies](https://mudpie.ai/companies/) / [Cashew Labs: self-improving agent systems](https://mudpie.ai/companies/cashew-labs/)
Author: Ali Abouelatta (https://mudpie.ai/authors/ali-abouelatta/)
Published: 2026-09-19
Updated: 2026-09-19
Research type: Company profile
Method: Company and accelerator sources checked 2026-09-19. Product claims are attributed to their sources; this is research, not a hands-on product trial.

## What it does

Cashew Labs is building open-source infrastructure for recursive self-improvement: models, agents and interfaces that can change how they work. The [YC profile](https://www.ycombinator.com/companies/cashew-labs) describes self-modifying interfaces and LLMs that post-train themselves; the [current homepage](https://cashew-labs.com/) names Libretto, a browser/toolkit for coding-agent automations, and Halo, a desktop workspace for custom AI interfaces.

The fit is a technical team experimenting with agent interfaces, browser automation or model-improvement loops. Cashew is not a conventional application with a fixed workflow. It is a developer/research infrastructure bet whose value depends on safe self-modification, reproducibility and the ability to understand what changed.

## Why I’d look closer

The open-source surface is the strongest reason to look. Libretto is aimed at agents that inspect websites, build automations and maintain integrations; Halo is positioned as a workspace where agents can build custom interfaces. That gives an evaluator tangible artifacts to inspect instead of only a research thesis.

The [founder biographies](https://www.ycombinator.com/companies/cashew-labs) describe Michael Kronovet as a former Palantir ML engineer who led State Department work, and Tanishq Kancharla as a product engineer and co-founder of Libretto. The same official page carries an older Saffron Health launch; I am not importing those healthcare claims into the current Cashew profile because the current homepage clearly presents the self-improving-systems identity.

## What I’d ask

How are self-modifying changes versioned and rolled back, what is the evaluation gate before a new behavior ships, and how are browser permissions isolated? I’d run a small agent against a test site, inspect generated artifacts and traceability, and compare the result to a static baseline. Pricing and a full hosted-product plan were not exposed in the sources checked.

## My editorial take

Cashew Labs is interesting for builders who want the agent and interface to evolve together. The open-source projects make that thesis inspectable. The decision should turn on reproducibility and safe iteration, not on the promise that systems can rewrite themselves.

## Quick facts

| Field | Sourced detail |
|---|---|
| Buyer fit | AI infrastructure, agent and browser-automation builders |
| Current projects | Libretto and Halo; open-source surface |
| Founder context | Palantir/State Department ML and product-engineering backgrounds |
| Public pricing | Not exposed in the sources checked |

## Sources checked

Checked 2026-09-19.

| Source | Used for |
|---|---|
| [YC company profile](https://www.ycombinator.com/companies/cashew-labs) | Product, founders and older launch distinction |
| [Cashew Labs homepage](https://cashew-labs.com/) | Current open-source project identity |

## Cohort context

Cashew Labs is listed in Spring 2025. In our 2026-09-18 directory snapshot, 97 of 143 listed companies in that cohort have YC’s primary industry label B2B (67.8%). This is a current-directory comparison, not an original intake count or a performance ranking. [Nine-cohort dataset](https://mudpie.ai/research/yc-cohorts-2026-09-19.json).

## Public website snapshot

Observed 2026-09-19T16:15:34.011Z in raw homepage HTML. This records visible metadata and advertised links, not agent execution or product quality.

| Signal | Homepage observation |
| --- | --- |
| Product description metadata | Observed |
| Canonical link | Not observed in this response |
| H1 or H2 heading | Observed |
| Typed structured data | Not observed in this response |
| Docs/developer link | Not observed in this response |
| Pricing link | Not observed in this response |
| llms.txt link | Not observed in this response |
| Markdown alternate | Not observed in this response |

[Public observations](https://mudpie.ai/research/yc-homepage-links-2026-09-19.json) · [Collection method](https://mudpie.ai/research/yc-homepage-methods/README.md). Missing links here do not establish that a capability or file is absent elsewhere.


## Author disclosure

I cofound Lazyweb and publish Mudpie. This is an owner-written publication, not an independent testing organization. Research notes distinguish observations, sourced reporting and editorial judgment.