støvetøp retweeted
running out of claude credits?
you could use less
you could use less
I've been using Adage for the past few weeks and I can notice the time and tokens its saved me
open source: github.com/stevenvaughanb…
npmjs.com/settings/adage…
I went to college with dev, best friend I could ever ask for, smartest & chillest dude at the function
check it out fr
3
3
22
1,435
Your agent shouldn’t pay to learn the same lesson twice. 🏁
Adage turns agent history into shortcuts.
40% fewer tokens. 22% fewer model calls. One advisory benchmark.
⭐ Star it. Help build the next shortcut. github.com/stevenvaughanb…
Adage turns agent history into shortcuts.
40% fewer tokens. 22% fewer model calls. One advisory benchmark.
⭐ Star it. Help build the next shortcut. github.com/stevenvaughanb…
0
0
1
46
You already paid for this lesson.
Every session, my coding agent rediscovered the test command that works and the step that follows a schema change. So I built Adage: an experiment that learns those from the history already on your machine and hands them back as reflexes.
8
Every session, my coding agent rediscovered the test command that works and the step that follows a schema change. So I built Adage: an experiment that learns those from the history already on your machine and hands them back as reflexes.
8
2
1
2
12,644
Nothing leaves your machine. Learning and matching run on your CPU, hooks make no network calls, and transcripts never go to a model or service.
Guidance by default. Running commands is opt-in with per-action approval. ADAGE_DISABLE=1 turns it all off.
Guidance by default. Running commands is opt-in with per-action approval. ADAGE_DISABLE=1 turns it all off.
1
0
1
20
The part I can't test alone: repos with strong conventions should produce the best reflexes.
If you have one, npm i -g @adage-sh/cli && adage init, then tell me what it learned, or didn't. Issues help me fix it. Stars help the next person find it.
If you have one, npm i -g @adage-sh/cli && adage init, then tell me what it learned, or didn't. Issues help me fix it. Stars help the next person find it.
0
0
1
14
3/ The numbers, honestly: 48% fewer model requests and 22% lower API cost across 6 sandbox tasks run twice on Sonnet 5.5. A held-out run on Haiku 4.5 got 22% fewer requests.
github.com/stevenvaughanb…
github.com/stevenvaughanb…
1
0
0
6
5/ The part I can't test alone: repos with strong conventions should produce the best reflexes.
If you have one: npm i -g @adage-sh/cli && adage init, then tell me what it learned, or didn't. Issues help me fix it. Stars help the next person find it.
If you have one: npm i -g @adage-sh/cli && adage init, then tell me what it learned, or didn't. Issues help me fix it. Stars help the next person find it.
0
0
0
6
støvetøp retweeted
I've been using Adage for the past few weeks and I can notice the time and tokens its saved me
open source: github.com/stevenvaughanb…
npmjs.com/settings/adage…
I went to college with dev, best friend I could ever ask for, smartest & chillest dude at the function
check it out fr
open source: github.com/stevenvaughanb…
npmjs.com/settings/adage…
I went to college with dev, best friend I could ever ask for, smartest & chillest dude at the function
check it out fr
My guy built a directed acyclic graph engine that turns coding agent traces into reflexes
Early numbers, Haiku 4.5, held-out tasks:
🛣️ 20 → 2 median turns before first edit
$0.135 → $0.053 per run (~60% cheaper)
🏃🏼♂️ <7 ms per decision, on a CPU, no model call
3
1
16
2,859


