Loa
This app is an open source coding agent named Loa designed to run against self-hosted LLM environments such as Ollama, targeting non-cloud-sized models (e.g. a 27b model). It decomposes tasks into very simple single steps that smaller models can solve, dynamically mutates and restructures its execution plan as new context arrives, and maintains a layered memory/context system combining chat history, step results, an action backlog, reasoning explanations, and a working memory of facts and decisions. It provides a full observability UI showing memory contents, composed context, plan dependencies, artifacts, and an audit log of all executed actions, accepts mid-execution steering messages that can alter or mutate the plan, and includes error-correction and self-reflection logic (e.g. JSON repair fallbacks and hypothesis-driven retries or plan changes). It deliberately trades execution time (many inferences) for reasoning and result quality, aiming to be a hands-off agent that is briefed once and completes the task without constant reprompting, and can also serve as an analysis and discussion partner. A stable beta is planned for release.
What it addresses
Existing coding agents target cloud-sized models and try to solve whole tasks in a handful of inferences, overloading the model and often requiring users to reprompt repeatedly until something is correctly solved; smaller self-hosted models struggle with monolithic multi-step tasks.
Shared over time
Each entry reflects what was shared on that date. Details may have changed since then.
Loa
This app is an open source coding agent named Loa designed to run against self-hosted LLM environments such as Ollama, targeting non-cloud-sized models (e.g. a 27b model). It decomposes tasks into very simple single steps that smaller models can solve, dynamically mutates and restructures its execution plan as new context arrives, and maintains a layered memory/context system combining chat history, step results, an action backlog, reasoning explanations, and a working memory of facts and decisions. It provides a full observability UI showing memory contents, composed context, plan dependencies, artifacts, and an audit log of all executed actions, accepts mid-execution steering messages that can alter or mutate the plan, and includes error-correction and self-reflection logic (e.g. JSON repair fallbacks and hypothesis-driven retries or plan changes). It deliberately trades execution time (many inferences) for reasoning and result quality, aiming to be a hands-off agent that is briefed once and completes the task without constant reprompting, and can also serve as an analysis and discussion partner. A stable beta is planned for release.
Problem: Existing coding agents target cloud-sized models and try to solve whole tasks in a handful of inferences, overloading the model and often requiring users to reprompt repeatedly until something is correctly solved; smaller self-hosted models struggle with monolithic multi-step tasks.
Audience: Developers who want a hands-off coding agent running on self-hosted models
Posted by @voodooEntity · Original comment ↗
“Ill wrap it up here and say : the Agents name is "Loa" and i probably will post it on hackernews as soon i got a relatively stable beta to release.”