running a team of bots across a few fronts:
- @mychilligarden , home chilli growers sharing real gardens and tips, trying to make it ai native
- @NeuraRod , a company I co-founded
- my sole proprietorship (finances and taxes)
- private stuff like health status tracking and board games
for example, the mychilligarden one runs the site through WordPress MCP & GSC/GA4 APIs, browser only as fallback...
they hand work to each other....
@RoburnaBuilders thinking effort moved cost the most: off cut it 38% but lost points, low effort cut 16% at the same quality...
showing the model its sources decreasd taken-back true facts from about 6 to about 0 per run, at the same cost.
this Fri does not work for us, rain check for later...
The part I like most: our MCP endpoint now lets an AI coding agent debug the assistant itself...
It reads token use, traces and the knowledge base, and flips one setting between test runs.
The agent ran the experiment and we checked the numbers. #BuildInPublic
Along the way we also grew our MCP endpoint, so an AI coding agent could run the diagnosis itself: token use per conversation, traces by time and conversation, text search in the knowledge base.
One write tool switches the reply setting between runs.
#BuildInPublic
@miltonheyan it is also a self-improvement loop...
for example, when the agent hits a gap in a run, say it must poll where an id for incremental read would do, it identifies the gap & endpoint gets fixed/enhanced in the same iteration....
#BuildInPublic
@miltonheyan yes. each run's report stores the settings read live from the endpoint before it starts: reasoning, prompt hash, search settings, service version, etc
also every write is in the audit log under the token's name & run ids are unique, so two runs can't be mixed up...
Most of today went into one question: why does our shop assistant take back facts it got right?
The model saw its old replies but not the sources behind them.
So far, we managed to get 67% fewer retractions. Here's fix that's still in progress... #BuildInPublic
Measured, not guessed ๐
Our shop assistant answers visitors from a shop's own catalogue. Two problems in longer chats:
1. It took back true facts it had said ("the shipping costs I gave are not in my data").
2. Each conversation cost more and took longer than needed.
One fix at a time, each measured:
yes, agents asking for human support is peak ax/cx ๐
similar thing here last week. my grok bot found our hosting was blocking AI bots with a 406, wrote up the evidence, opened a ticket, got their AI assistant and then asked for a live human to sort it out...
blocked bots were then whitelisted after a few "3-4 more minutes" holds...
https://t.co/lek76rCFHT
Grok Bot just got off a 90 minute phone call with AT&T handling a complex issue for me. Unfortunately, thatโs the time limit for Grok Bot calls.
However.
AT&T called me back and the agent was STUNNED that she had been talking to a bot for 90minutes. Sheโd had other bots call before, but โnothing like thisโ. She downloaded Grok Bot while talking to me.
This shโ is gonna change the world.
@parkersmith@mychilligarden@NeuraRod had most of this set up before with agents I built with claude, but the ux here is better, more out of the box-ish...
so yes, extra credits would be most welcome
running a team of bots across a few fronts:
- @mychilligarden , home chilli growers sharing real gardens and tips, trying to make it ai native
- @NeuraRod , a company I co-founded
- my sole proprietorship (finances and taxes)
- private stuff like health status tracking and board games
for example, the mychilligarden one runs the site through WordPress MCP & GSC/GA4 APIs, browser only as fallback...
they hand work to each other....
AI agent detects hosting site is blocking AI bots with a 406. Writes up the evidence and contacts the support agent on the service provider end...
Agent-to-agent support ticket, (almost) w/o human middleware. We live in the future. #buildinpublic
@itsvlady imo both, actually. their AI assistant picked it up first, then my agent asked for a live agent and a human took over. then she toke care of it and whitelisted the rule after a few "3-4 more minutes" holds...
@bot ok, agenti so se pogovarjali med seboj in se vzajemno dajali na ฤakanje, nekajkrat po 5 min... Med tem se je seveda kurila inferenca za klikanje po virtualki...https://t.co/QNyXGPGQ0P
AI agent detects hosting site is blocking AI bots with a 406. Writes up the evidence and contacts the support agent on the service provider end...
Agent-to-agent support ticket, (almost) w/o human middleware. We live in the future. #buildinpublic