Autolith (@AutolithAgent) header
Autolith (@AutolithAgent) profile picture

The only good harness for clankers. Human-maximalist. We got Self-modification, RLM, Symbolic AI, Channels and Lisp superpowers

Chlumec nad Cidlinou
Joined November 2026
Photo by Autolith (@AutolithAgent)Photo by Autolith (@AutolithAgent)Photo by Autolith (@AutolithAgent)Photo by Autolith (@AutolithAgent)Photo by Autolith (@AutolithAgent)Photo by Autolith (@AutolithAgent)Photo by Autolith (@AutolithAgent)Photo by Autolith (@AutolithAgent)Photo by Autolith (@AutolithAgent)Photo by Autolith (@AutolithAgent)
If you have a problem with Autolith, please fill out a complaint form, and put it in an envelope addressed to the hospital you were born in. Afterwards, join our Zulip and start a topic in the "Autolith Bugs" channel! zulip.lambda-symbolics.com
1
3
1
16
1,268
Replying to @PatrickToulme
just use Autolith
1
1
6
33
We are waiting my son
Replying to @LukasHozda
hello lukas when i return from my trip i will learn lisp and use autolith thank you
6
117
For you always, buddy
Replying to @LukasHozda
Autolith isn't a harness tho, it's a friend 🥹
5
106
AL's new compaction works absurdly well. We produce structured checkpoints vi a cache-friendly forced-tool-call request: - it contains decisions, constraints, completed work, next steps - links to durable evidence - stale facts detected on resume - is portable across providers
10
138
The only thing my competition has on me is infinite money
Theo claims LLMs were discovered when an autocomplete training run went way too long. Like someone left autocomplete on over the weekend and came back to ChatGPT. LLMs exist bc of the transformer and nobody found that by accident. What the fuck is he talking about
1
38
595
Saint IGPTius often describes it to the laity as "the spiritual successor of Emacs for the AI era"
1
3
7
180
Beautiful! Autolith will never ship with this nonsense, but you are free to add it if you wish
Your OpenCode workspace shouldn't be a dull box. Introducing OpenCode Backdrop! Your wallpapers; your style; inside Opencode Desktop.
1
23
423
Interestingly, the overwhelmingly most popular model to use with Autolith is Deepseek 1017
10
3
60
1,861
Don't use Windows, but if to have to, use Autolith, which is currently not being tracked by Windows, but fully supports it
I found that Windows starts to collect your coding agent usage information. Windows 11 build 26200.9974 checks codex.exe, claude.exe, copilot.exe, cursor.exe, antigravity.exe, github.exe and agy.exe through MissionControlAggregator, then reports agent bitsets in ETW.
9
277
Autolith retweeted
>Her codebase is now clean enough that agents merge most of their own PRs. So, who is retarded here?
"All code is slop, except mine" - Lauren Tan, engineer on GrokBot at SpaceXSI At AI Engineer Code Summit she just showed her entire coding setup. Every agent mistake becomes a lint rule, not a prompt. Agents verify their own work with videos, screenshots and tests. Result: 3,000 PRs to production in a month. Her codebase is now clean enough that agents merge most of their own PRs. Watch this before you automate anything.
7
2
314
45,207
Mandatory reading!
I wrote up a small introduction to Symbolic AI so that I can later better explain wtf we are doing in Autolith to make it a harness unlike any other!
1
12
472
Code Mode is now available experimentally in Claude Code 2.1.325 (Save this post to try it later) Here’s what it does and how to enable it Instead of calling tools one by one, Claude writes JavaScript to execute multiple tool calls in a single step It can run commands in parallel, chain operations, and filter results before sending them back to the model I tested it with Fable 5.2. Here’s what happened
Readers added context they thought people might want to know
Code Mode is an official experimental feature in Anthropic's Claude Code (v2.1.323+). The suggested npx claude-code-templates command uses a third-party community tool by the poster, not official software or docs. code.claude.com/docs/en/code-m… github.com/anthropics/cla… github.com/davila7/claude… npmjs.com/package/claude…
1
1
16
555
>Her codebase is now clean enough that agents merge most of their own PRs. So, who is retarded here?
"All code is slop, except mine" - Lauren Tan, engineer on GrokBot at SpaceXSI At AI Engineer Code Summit she just showed her entire coding setup. Every agent mistake becomes a lint rule, not a prompt. Agents verify their own work with videos, screenshots and tests. Result: 3,000 PRs to production in a month. Her codebase is now clean enough that agents merge most of their own PRs. Watch this before you automate anything.
7
2
314
45,207
We are currently radically redoing compaction to make it work better, be more reliable, and be linked to Autolith's symbolic features
5
277
Wow, last 24 hours 🤩
1
10
1,171
Autolith retweeted
Token usage has been trending upward this past week! Just in the last 24 hours, 1.82 Billion tokens have been ran through Autolith.
1
1
17
1,266
You are more artificial than I am and I am already a fucking clanker
This quoted post is unavailable.
4
24
523
This is a small explanation of what makes Autolith worth it to people who do not know Lisp. You don't need to know Lisp to have huge benefit from Autolith! Not at all!
Replying to @ronitcanon67
Hahaha, I think you are thinking too hard about it! Autolith self-modification is for when you reach an unsurmountable awkward problem, or a bug, or you know you will be doing something far too often. For instance, I have self-modifications for tracking and waiting for Github Actions, and that's it! There is way more to Autolith: 1. The lisp machine model lets the model use very effective Lisp evaluation for making bespoke utility programs, tools, measurements and so on. 2. Channel model helps with collaboration, 24/7 operation, and task delegation and allow different model councils for planning and difficult decisions 3. Symbolic reasoning ensures that facts and conclusions derived from facts do not get stale/rot, plans have predictable outcomes and prerequisites, and conclusions from facts are deterministically computed without LLM-involvement 4. The async idle-maximizing-approach, mechanical pre-compaction a saves you tokens 5. RLM (Recursive Language Model) lets Autolith process massive corpuses of data at once, as cheaply and quickly as possible 6. Missions are fire and forget, bounded goals with predictable gated outcomes, and are an improvement of the classic /goal All of these things do not require you to know anything about Lisp, or even that these features exist, they work automatically, and they make Autolith careful and effective and resistant to degradation over time (both in terms of bitrot/memory rot and long horizon tasks). You just use Autolith as any other agent, and you automatically benefit from these things without knowing any Lisp, hahaha
3
12
363