Lambda Symbolics releases Autolith programming agent
Lambda Symbolics has released Autolith v0.35.0, a terminal-based programming agent with a live Common Lisp runtime that lets developers inspect, extend, and recover their AI assistant.

Autolith v0.35.0 runs directly inside a user's repository on a terminal, supporting platforms including Linux x86-64, Linux aarch64, macOS arm64, FreeBSD x86-64, NetBSD x86-64, and OpenBSD x86-64. The binary release packages its own Steel Bank Common Lisp (SBCL) 2.6.6 runtime, Lisp dependencies, and native helpers. Rather than acting as a simple wrapper, the agent operates within a single process containing the provider client, terminal interface, tool registry, and conversation state. It can execute model-generated code with user privileges, though developers can also run it in an immutable mode to restrict evaluation and mutation tools.
The agent integrates with several AI backends, including ChatGPT Codex subscriptions, Grok subscriptions, Fireworks AI, the Anthropic API, and OpenCode. Other OpenAI-compatible endpoints can be registered manually. To handle codebases larger than an LLM's context window, Autolith treats oversized corpora as an environment rather than a prompt. For example, in a demonstration using the gpt-5.6-terra model, the agent successfully analyzed a 3.1-megabyte file containing 121 concatenated source files. Operating under a budget of 32 calls and 400,000 tokens at depth two, the model identified 83 condition classes across 14 subsystems in just over four and a half minutes, without the corpus ever entering the prompt directly.
A key feature of Autolith is its ability to modify its own running image. Developers can inspect and redefine functions, methods, and classes on the fly. These changes are recorded in an append-only mutation journal and can be saved as private image commits, such as commits 4048480f and 9eee3ae4, which are stored in a separate private Git history. If a model-generated change crashes the active process, Autolith writes a crash capsule, like capsule efa6aa16, and boots a pristine recovery image. This separate image inspects the crash and restores the conversation state from a clean, committed source without loading the damaged core.
This is our own summary of reporting by Hacker News


