5 free GitHub repos I'm saving for my AI agents. Testing one this week:
1. OpenCode: coding agent
2. Superpowers: one method for Claude Code, Codex, Cursor
3. anthropics/skills: starter skills
4. Langfuse: see every step your agent took
5. Learn Harness Engineering: free course
@OpenAI provenance matters even at my size. the notes claude leaves for my bot team say which model and effort level wrote them, so i know whose work i'm reading.
with a distilled copy that label stops meaning much fr.
@omarsar0@AgentMemoryL my preferences sit in a plain written file. one line says reel scripts come to me as spoken lines in one block, with a blank line at each shot change.
has anyone scored a plain file like that on the care category?
@AravSrinivas i'd push back on evidence being the gate. the bot that checks my whole team every monday can only propose changes, it's not allowed to touch another bot.
so every fix still waits on someone saying yes.
@claudeai mine's the dullest one going. just after 5am claude checks my mac has checked in within the last 12 hours and that the backup pushed.
thursday's whole report was two words. all good.
@emollick my bots get their model from a written rule, and my own rules file admits that's only an instruction, since nobody can swap the engine underneath directly.
so if a point release ever went weird, i'd notice it in the work first.
@omarsar0@Netlify in my setup one bot owns each build and does its own testing and fixing in one claude code run, then reports back once.
the part still on me is saying yes before anything goes live.
@OpenAI@AMD love that the models improved designs AMD engineers had already spent years optimising.
my bots run a tiny version of that. each one is told to read the same rules file before it writes anything for me.
how much of the gain came from the engineers' own past work?
@elonmusk@bot i'm moving six of my scheduled jobs from claude to grok bot, and each one only switches over after a clean week.
on tuesday the ai news job landed 24 minutes after its target time and still counted as a yes.
will grok bot ever warn me before it runs late?
@mattshumer_ my morning news bot said it was a quiet night for openai.
by this afternoon my whole feed was gpt 6.2 reviews, so much for quiet.
did any of the reviewers say what real jobs they tested it on?
@DanKornas my bots share a brain that lives in a github repo, so every claude chat reads the same rules and notes.
grok bot writes each finished task there as a dated file for claude to pick up.
how does puppyone handle two agents editing the same file at once?
@bindureddy for my bots, agentic means one boring thing: did it follow the rules file.
every morning claude marks what my grok bot ran the day before. yesterday 10 ran, 7 passed and 3 needed a fix.
do your agentic tests score rule following, or just whether the task got done?
@rileybrown i had my automation bots on fable 5.1 for about a day on 30 oct, then moved them to opus 5.5 for everything on 31 oct.
what lands in your 15% fable pile, planning or the quick stuff?
@gdb security for a one person setup is mostly boring stuff.
my backup job skips any file that looks like a secret, and this morning it was holding back 12.
will gpt 6 cyber ever look at small setups like that, or is it mostly big codebases?
@AravSrinivas i split models by job instead of running them side by side.
fable 5.2 writes the plan, i approve it, then opus 5.5 does the work and proves it.
when two models in multiagent disagree, who breaks the tie?
@GeminiApp the hard part with recurring tasks is the night one quietly doesn't run.
on monday night my grok bot's nightly sync didn't land, and my morning brief flagged all 17 monday tasks as missing.
does gemini tell you when a scheduled task fails, or do you find out later?
@chaibytesai how mine hand off: each step writes one file at a set time. the next bot works from whatever exists, so one missing step doesn't stall the rest
which of your two breaks hurt more?
@DanKornas storage was never my bots' problem. stale rules were
on 31 oct i set their model rule at 3pm and had to change it by 3:26
now rules get a date and the newest one wins
how does agentdb know a memory is out of date?
@GoogleDeepMind@antigravity question from the small business end
does lab grade reasoning hold up on boring multi step admin, or only on clean problems with a right answer?