Google Tokenade, the Lean Harness for your agent
Joined October 2021
- Tweets899
- Following167
- Followers98
- Likes341
Your AI runs out of room re-reading everything it already saw, at every step. One big file read early is paid again at each step after it.
Count what it re-reads before you touch the model.
Restart with a short summary, or cut the heavy output.
Each agent turn re-sends the whole session, including every dead end.
On my open benchmark of token-saving tools, long sessions cost 39% less with Tokenade, on average:
pi-infected.github.io/token-…
Tunguz forecasts inference passing the database market, about $350B by 2027. Still a forecast.
When usage costs more than the license, every token your agents send comes out of your margin.
Measure what one client project costs in tokens before you quote the next.
Effective token prices fell 41%, and usage rose 50% since July (Ramp AI Index, Sept 2026).
Waiting for a cut won't shrink the bill, since usage outruns price.
Your agent rereads its whole context every turn. Volume per task is the lever.
Pi's bash tool shows the model at most 2,000 lines of output. With Codemode, the agent scripts its tool calls and filters the results before the model reads them.
The model should read what it needs, not everything the tool printed.
My paid model didn't hold. Most people never see the token bill, so nobody pays to save it.
Tokenade 2.0 is free and open source, MIT OR Apache-2.0.
Two commands: npm install -g @tokenade/cli, then tokenade install. The code is readable on GitHub.
github.com/pi-infected/token…
A bigger context window does not make an agent cheaper.
DeepSeek V4.1 Flash 1017 shipped October 17 with 1,048,576 tokens of context. Every call resends the history, so the re-reads drive the bill.
Have you plugged it into an agent yet?
An agent can start paid cloud services overnight. A budget alert arrives after the invoice. A hard cap stops the spend first. AWS began rolling out spend caps on 16 September, Google Cloud has had them since July. Set one before your agent runs.
13M tokens saved in a month, worth €178.53. Not my benchmark: an early Tokenade user read it off their own dashboard. Their words: "Very simple to install and use."
Check your own counter. It beats my charts.
"Tokenade doesn't sell dreams." A user sent me that.
Another, who runs heavy coding agent tasks: "a solid 30 to 50% depending on the case."
Their sessions. Not mine, not yours. Run your longest session through it and read the bill.
tokenade.net
Prospects I found in comment sections replied 22.2% of the time. From a people search, 13.2% (my outreach, Oct 2, 2026).
A profile can match every filter and belong to someone who no longer logs in. A fresh comment can't.
Comments only, now.
Your Claude Code limit fills up with tokens your agent already read.
Try Tokenade on your next long one. I benchmarked it against other token saving tools, method public: pi-infected.github.io/token-…
Step 5 Preview comes from StepFun, one of China's six "AI tigers" alongside Moonshot and MiniMax, backed by Tencent.
$1 per million input tokens. My read: Sonnet-class for 3 to 5x less.
Running agents unattended? Test it this week.
One agent on your own laptop: if it misbehaves, you shrug. Two hundred seats running agents all day: something will go wrong, only the date is open. The first question after is "what did it touch?" Most AI tools keep no log to answer it.
Quotas and a cheaper model cut the AI bill on paper. Your engineers pay it back in waiting.
Cut what the agent rereads each turn instead. Tokenade (I built it) cuts token costs 39% on average on long sessions: pi-infected.github.io/token-…
Checking whether a date is in the past should not cost a single token.
In the agents I ship, counting, filtering and format checks never reach the model. They're code, with a test, and they answer the same way every run.
The model gets the calls that need judgment.