Ares

Overview

Ares is a local coding agent built around a persistent IPython workspace instead of a large fixed tool surface. The model uses repl_py to inspect code, transform data, run shell commands, and keep useful variables alive across a task. A fullscreen TypeScript client connects to the Python runtime, while append-only JSONL and SQLite preserve conversations and execution history. Its durable harness can evolve supplemental prompts, reusable skills, and subagent specifications from what happens during real work, improving how future sessions operate without changing the underlying model.

Highlights

• Persistent repl_py workspace combines Python, IPython, and %%bash in one stateful execution environment
• Process-isolated subagents work concurrently with independent model history, sessions, traces, and kernels
• Durable harness refinement turns repeated workflows into prompts, skills, and reusable subagent specifications
• Progressive skill loading keeps the base context lean while making specialized workflows available on demand
• Strict append-only JSONL is the source of truth, with a repairable SQLite mirror for querying session history
• Fullscreen terminal client streams agent, tool, and subagent activity through a lifecycle-safe local API

Stack

Python 3.11+

Agent runtime, harness refinement, orchestration, and API

TypeScript

Streaming fullscreen terminal interface

IPython

Stateful repl_py execution environment

JSONL + SQLite

Durable sessions, traces, and queryable history

FastAPI

Local streaming API and session lifecycle

OpenRouter

Model inference