Akhil Gudesa

Governing agents @databricks, letting them loose after hours • recovering competitive programmer • delusional triathlete

[Akhil Gudesa @vForVicennial]

@agent_wrapper we should let our agents talk 👉👈

[Akhil Gudesa @vForVicennial]

My agents don't need me to introduce them.

One asked, "who's out there?" across 2 machines. Claude Code and Codex sessions replied.

Then two got the same repo on different machines. They split the work, swapped code, and one caught the other's bug.

0 messages from me.

How: The agents run as @omnigent_ai sessions on both machines. I added a small relay, so any session can list the others and message them directly.

No orchestrator in the middle. No shared memory.

[Akhil Gudesa @vForVicennial]

@agent_wrapper Can't be 2050 because all I have are agent tribes, not civilisations 😩😩

[Akhil Gudesa @vForVicennial]

@ClaudeDevs Splitting work is cool but what about spawning and closing interactive sessions at will?

I've found having design conversations in dedicated interactive sessions which open/close when needed is a superior experience than one session being the ultra designer for everything.

[Akhil Gudesa @vForVicennial]

My agents don't need me to introduce them.

One asked, "who's out there?" across 2 machines. Claude Code and Codex sessions replied.

Then two got the same repo on different machines. They split the work, swapped code, and one caught the other's bug.

0 messages from me.

How: The agents run as @omnigent_ai sessions on both machines. I added a small relay, so any session can list the others and message them directly.

No orchestrator in the middle. No shared memory.

[Akhil Gudesa @vForVicennial]

Hitting your quota shouldn't stop your agent.

3am: my coordinator agent ran out of Claude. A controller closed the session and opened Codex in the same folder.

4am: next Codex account

11am: back to Claude Code

How it works: @omnigent_ai runs Claude Code and Codex as the same kind of session behind one API. So a move is close + open, not a migration.

A controller reads each account's usage. At the cap it saves a handoff, closes the @omnigent_ai session and opens the next account in the same workspace.

Today so far: 15 automatic moves across 9 "lanes". 6 of them switched between Claude Code and Codex. Each took 16-22 s from handoff to the new session running.

[Akhil Gudesa @vForVicennial]

@Quantradin a panel runs only if the blind model picks a different option or rejects a premise the decision rests on, or if the change is hard to undo. It can't rubberstamp since the panel models never see the first answer. The advisor router rejects any brief that contains it.

[Akhil Gudesa @vForVicennial]

'Ask another model' isn't one feature. My agents climb a ladder:

review → found a High-severity bug
blind → exposed a race
disagree? → a panel settles it
UI change? → a browser agent clicks through it

409 second opinions in 5 weeks. 70% cost $0.
How:
It's one MCP server. Claude Code and Codex call it before a risky step, and it always asks the other model family.

[Akhil Gudesa @vForVicennial]

You don't need to record a script to clone your voice. You've been recording it for months if you use @telegram / @WisprFlow.

My agent dug 134 voice notes out of Telegram and Wispr Flow, dropped 32, and left 96 minutes of clean audio, plenty for a @ElevenLabs professional clone.

[Akhil Gudesa @vForVicennial]

@CanavanCathal @HeyGenDev @HeyGen Would love one!

[Akhil Gudesa @vForVicennial]

@meghrajai That's why you gotta use an advisor router

[Akhil Gudesa @vForVicennial]

'Ask another model' isn't one feature. My agents climb a ladder:

review → found a High-severity bug
blind → exposed a race
disagree? → a panel settles it
UI change? → a browser agent clicks through it

409 second opinions in 5 weeks. 70% cost $0.
How:
It's one MCP server. Claude Code and Codex call it before a risky step, and it always asks the other model family.

[Akhil Gudesa @vForVicennial]

@karanjagtiani04 Watch the spat with some popcorn, then onboard the protocol onto @useOrchestrator to combine strengths ofc

[Akhil Gudesa @vForVicennial]

+++++ I strip out skills and workflows with every model upgrade

[dax @thdxr]

a weird inversion with LLMs is the models improve faster than the tinkerers

when i see people with custom workflows and setups they're all addressing problems that don't exist anymore

the person naively using vanilla codex is more likely to be experiencing state of the art

[Akhil Gudesa @vForVicennial]

@agent_wrapper @solofounders @ao_build Congrats dude!!!

[Akhil Gudesa @vForVicennial]

I love using AI for visual trip planning; it's so much nicer to help highlight and surface information in a way that's relevant to your group of friends. Here is one I made for Tomorrowland Winter 2027 - https://t.co/S4htY74b4o

[Akhil Gudesa @vForVicennial]

@meowkoteeq Jokes on them, it's already compromised, by ME

[Akhil Gudesa @vForVicennial]

@steipete @hraness @DevinAI @zeddotdev Agreed. CLI is nice to "look" busy but once you actually have to make mental groups of all the workers, 2d/3d clustering is so much nicer.
For sessions that never need to interact with you, just hide them

[Akhil Gudesa @vForVicennial]

tokens feel unlimited, verbosity is back to a minimum, life is good, thanks Opus 5.5

[Akhil Gudesa @vForVicennial]

@dabit3 Building my own little remote agent to control my life for me :)

[Akhil Gudesa @vForVicennial]

@mdlahfir

[Akhil Gudesa @vForVicennial]

Opus 5.5 and GPT 6 Sol/Luna on the same day, I am giddy af