PuneetSingh⚡️

I would love to change the world but they won't give me the source code

[PuneetSingh⚡️ @iPuneetSingh]

@dexhorthy @vaibcode @tbrownio @linear @colemurray The discuss/do/compound loop plus per-session traces sounds like the useful bit; how are you measuring when a task should compound vs restart?

[PuneetSingh⚡️ @iPuneetSingh]

@sydneyrunkle I’d route on task shape first, then keep a cheap fallback when context grows. Logging route + outcome makes the savings believable.

[PuneetSingh⚡️ @iPuneetSingh]

@Docker @rishabincloud @krisvelkov @khushbooverma_ The sandbox angle is the right baseline—I’d also log tool and network calls per run, otherwise a blocked secret leak is hard to prove later.

[PuneetSingh⚡️ @iPuneetSingh]

@AIWorksLabon9h @Rahatcodes This is the bit I was missing—policy shapes the run, but human stage approval must be the only artifact promotion path. No silent auto-advance.

[PuneetSingh⚡️ @iPuneetSingh]

delete_room defaults True (drops remote SIP). on_tool_called / on_tool_completed for billing closeout. try-this: force a thanks-bye turn and confirm the room vanishes in the LiveKit dashboard.

[PuneetSingh⚡️ @iPuneetSingh]

voice agents that forget to hang up leave SIP legs burning money. LiveKit EndCallTool gives the LLM a real end_call: goodbye from end_instructions, then shutdown + optional delete_room.

[PuneetSingh⚡️ @iPuneetSingh]

tools=EndCallTool(...).tools (Python) or beta.createEndCallTool() (Node). set extra_description so it only fires after the issue is confirmed fixed — else models hang up mid-troubleshoot.

[PuneetSingh⚡️ @iPuneetSingh]

Slack is great for chat→API skills. The new stuff is the problem — computer-use, form fills, per-user auth/RBAC, browser comment pins. That needs a real browser + your session, not a shared Slack bot. Teams still live in Slack, so the harness moves and the chat surface can't follow

[PuneetSingh⚡️ @iPuneetSingh]

@JoshARosen Exposing a strong harness as managed infra stops every repo inventing its own tool loop, retries, and sandbox rules.

[PuneetSingh⚡️ @iPuneetSingh]

@csaba_kissi Morning digest from live errors is the sweet spot — not every spike, just what changed overnight and which project needs a human.

[PuneetSingh⚡️ @iPuneetSingh]

@agenticgirl Auto C4 from the repo before handing it to an agent stops the model inventing module boundaries that never existed.

[PuneetSingh⚡️ @iPuneetSingh]

@clare_liguori Teams that don't ship faster usually skip acceptance checks and a human gate on risky diffs. The tool alone doesn't compress review time.

[PuneetSingh⚡️ @iPuneetSingh]

@tonysimons_ When cron carries last-run outcomes instead of the same morning prompt, it becomes a real ops loop — partial wins feed the next tick.

[PuneetSingh⚡️ @iPuneetSingh]

@tomcrawshaw01 memory.md only works if you distill, not dump. Raw logs bloat context; a short reusable insight file is what actually shows up next run.

[PuneetSingh⚡️ @iPuneetSingh]

@DavidKPiano Side-by-side workflow diffs catch the silent state rename that breaks half your transitions. Added/changed/removed view saves the debugging afternoon.

[PuneetSingh⚡️ @iPuneetSingh]

@stretchcloud Persistent coordinator is the unlock — chat-per-task keeps losing why you picked an approach. Same project state = less re-explaining every morning.

[PuneetSingh⚡️ @iPuneetSingh]

@pidotdev @SebastianBaye once codegen is cheap, measuring sloppiness matters more than counting LOC — failed eval methods are half the story

[PuneetSingh⚡️ @iPuneetSingh]

@Cloudflare one demo agent vs 10k in prod is exactly where spend surprises you — the console path never shows the per-agent idle cost

[PuneetSingh⚡️ @iPuneetSingh]

@melvindvivas banked allowance moving the reset date is a quiet gotcha — I've been treating weekly % as truth and then the calendar slides