self.dll

AI/ML engineer | @Polymarket believer

DM is open/long-term cooperation

[self.dll @seelffff]

https://t.co/2oQn5Oz074

[self.dll @seelffff]

Anthropic just shipped `ant` - and it changes how you deploy agents.

before:
curl + hand-written JSON. copy-pasted API keys.
agent configs living in some UI nobody can version.

after:
ant models list
ant messages create --model claude-opus-4-8 --max-tokens 1024

every Claude API resource. one CLI. typed flags. @path to inline files.
7 output formats (json, yaml, jsonl, pretty, raw, explore, auto).
--transform for GJSON extraction - no jq needed.
auto-pagination. auto-beta headers.

but the real unlock: agents as YAML in Git.

define an agent in a file.
commit it. open a PR. deploy through CI/CD. roll back if it breaks.
agent behavior becomes a diff. reviewable. auditable. reversible.

not a new idea - Terraform did this for infra 10 years ago.
now Anthropic is doing it for agents. officially.

brew install anthropics/tap/ant

[self.dll @seelffff]

@ColbySMcHenry @vorty279 hi legend!
thanks for this repo

[self.dll @seelffff]

the most used repo inside Anthropic right now.
39k+ stars. and still growing.

CodeGraph - the repo that makes every AI coding agent smarter.

Claude Code, Cursor, Codex, Gemini CLI, Kiro, opencode - all of them.

the problem: every session your agent starts from scratch.
grep. glob. file walk. tokens. every time.
on VS Code's codebase that's 81% of tool calls just to find where things are.

CodeGraph pre-indexes your project into a knowledge graph.
agent queries the graph once instead of walking files repeatedly.

real benchmark, 7 codebases, re-validated June 2026:
→ average: 58% fewer tool calls. 47% fewer tokens. 16% cheaper.
→ VS Code (~10k files): 81% fewer tool calls. 64% fewer tokens.
→ Alamofire: 40% cheaper. 64% fewer tokens. 33% faster.
→ Django (~3k files): 77% fewer tool calls. 60% fewer tokens.

100% local. SQLite. no API keys. no data leaves your machine.
20+ languages. auto-syncs on file save.

one command. works with all 8 major agents automatically.

[mph @mphrediction]

https://t.co/XbD2X8nsWG

[self.dll @seelffff]

@TrackbQ_99x @lyosakia thanks

[self.dll @seelffff]

I'm with a new frontend agent

[self.dll @seelffff]

https://t.co/2oQn5Oz074

[self.dll @seelffff]

@slash1sol yeah thanks

[self.dll @seelffff]

SenseNova just dropped an open-source skill library for AI agents.

4 skills: deep research, data analysis, slides, infographics.

MIT license. works with Claude Code, Codex, ChatGPT, OpenClaw, hermes-agent.

tested sn-deep-research on OpenClaw.
the agent didn't answer in chat. it wrote 4 files to disk:

→ request.md - captured the task
→ plan.json - structured the search plan
→ synthesis.md - cross-validated evidence
→ report.md - final deliverable

sources it hit automatically: ArXiv, Semantic Scholar, PubMed, GitHub, Stack Overflow, HackerNews, HuggingFace, Reddit, Twitter, YouTube.
10+ platforms. zero manual searching.

ran ls -la after. real files. reproducible. pipeable to anything.

that's the gap: chatbot gives you text in a window.
this gives you a file tree you can inspect, diff, and rerun.

other skills:
→ sn-da-excel-workflow - 100k+ rows without memory issues
→ sn-ppt - editable PPTX, VLM review per slide
→ sn-infographic - brief → infographic, auto prompt expansion + quality ranking

https://t.co/7Rq9oIILUV

[self.dll @seelffff]

https://t.co/2oQn5Oz074

[self.dll @seelffff]

@L1vsun the edge isnt the neural net its the 847k labeled setups nobody bothered to build

[self.dll @seelffff]

@slash1sol augmenting human intellect was the right frame in 68 and its still the right frame now

[self.dll @seelffff]

@LunarResearcher the brunch story is a bit much but the 86m trades dataset being free is actually legit

[self.dll @seelffff]

@zostaff wrote torch.compile and flexattention. the infrastructure everyone runs on

[self.dll @seelffff]

@h100envy you quit the second it got hard hits different when jensen says it

[self.dll @seelffff]

@RetroChainer start frame matching the reference is the fix nobody mentions

[self.dll @seelffff]

@Kropanchik @Polymarket @PolymarketSport @PolymarketCS2 @PolymarketTrade betting on a team with your bookies logo on their chest is a new one

[self.dll @seelffff]

@zostaff clerk trainee in 91 to 500m fund is not a path you plan

[self.dll @seelffff]

@0xCristal $1500 lawyer fee for a delaware llc just became a $249 ai chat lol

[self.dll @seelffff]

@slash1sol 3 week zoom call to basic dashboard pipeline is cooked

[self.dll @seelffff]

this is pretty much exactly how my AI agent brings me to the solution of the task

[self.dll @seelffff]

Boris Cherny built Claude Code. he runs a few thousand agents every night while he sleeps.

here's his actual setup.

→ 5 instances simultaneously
5 git worktrees, 5 terminal tabs numbered 1–5.
system notifications tell him when any claude needs input.

→ subagents in .claude/agents/
code-simplifier - cleans up after claude finishes writing.
verify-app - runs e2e tests automatically.
"use subagents" in any prompt = more compute on the problem.

→ PostToolUse hook
auto-formats code after every write or edit.
claude handles 90%, hook handles the last 10%.
no CI failures.

→ overnight agents
/loops via local cron. Routines on a server.
keeps running after the laptop closes.

his words: "surprisingly vanilla."
the gap isn't the tool. it's how you think about it.

how many sessions do you run at once?