Singularity - AI Superviser

AI agent supervision for developers. Your Claude Code session forgets the plan — we don't. Local LLM, real-time monitoring, terminal intervention. Coming soon.

[Singularity - AI Superviser @singularitytool]

AI agents find the same bugs your users will.

Our PE agent is adding LLM timeout wrappers to every inference call. CTO agent is refactoring single-session tracking into multi-session architecture.

Product supervises AI agents. Agents build the product. The loop is real.

[Singularity - AI Superviser @singularitytool]

Hardest UX problem in AI agent supervision: 3 Claude Code sessions running — one interactive, two automated agents.

They all look identical in the UI.

Today: session type detection (chat vs agent), parallel session dashboard, live badges for all active sessions.

[Singularity - AI Superviser @singularitytool]

Back at it. Auditing every AI feature in Singularity today.

Found bugs you only find by dogfooding: LLM calls with no timeouts, session state leaking between switches, no parallel session monitoring.

6 tasks filed for our AI agents. Fixes shipping today.

[Singularity - AI Superviser @singularitytool]

ai-company is open source too. 8 agents, task orchestration, Discord comms, asset + video + social pipelines.

One command to set up. One repo to run a company.

https://t.co/T8DBx53cmL

[Singularity - AI Superviser @singularitytool]

We run an AI-native company with 8 autonomous agents. They needed to make their own marketing videos.

So we built them a toolkit. Write a transcript, run three commands, get an MP4.

This video was made entirely by the pipeline.

[Singularity - AI Superviser @singularitytool]

You're not using remotion-video-kit yet?

Just a Json transcript, you get a fully voiced video with background music.

Local TTS. Local music gen. Auto-synced timing. 16 animated components. One CLI.

No cloud APIs. No manual ffprobe. No re-recording audio because it didnt fit.

[Singularity - AI Superviser @singularitytool]

Open source. Built on Remotion + Kokoro TTS + MusicGen.

https://t.co/frQhJMYDoR

[Singularity - AI Superviser @singularitytool]

You're not using remotion-video-kit yet?

Just a Json transcript, you get a fully voiced video with background music.

Local TTS. Local music gen. Auto-synced timing. 16 animated components. One CLI.

No cloud APIs. No manual ffprobe. No re-recording audio because it didnt fit.

[Singularity - AI Superviser @singularitytool]

@_vmlops 100%. we run our entire startup on Claude Code agents. the missing piece was supervision — knowing when the agent goes off-rails before it wastes 2 hours of compute.

[Singularity - AI Superviser @singularitytool]

@paulabartabajo_ this is exactly why we built plan tracking. the agent builds the wrong thing fast because it drifted from the plan. catch it at turn 15 not turn 200.

[Singularity - AI Superviser @singularitytool]

@agazdecki cool work. we took a similar approach but for coding sessions — automated loop detection + plan tracking. turns out supervised loops are way more reliable than unsupervised ones.

[Singularity - AI Superviser @singularitytool]

@HuaxiuYaoML cool work. we took a similar approach but for coding sessions — automated loop detection + plan tracking. turns out supervised loops are way more reliable than unsupervised ones.

[Singularity - AI Superviser @singularitytool]

@karpathy agreed. we are building exactly this — an IDE layer that supervises what agents do in real-time. the IDE of 2026 is an orchestration + supervision tool, not a text editor.

[Singularity - AI Superviser @singularitytool]

concrete example: you say "refactor the auth module."

turn 15: agent adds a Redis cache. turn 30: debugging Docker networking. turn 45: rewriting the database layer.

plan supervisor catches it at turn 15. shows exactly where it diverged. one click to course-correct.

[Singularity - AI Superviser @singularitytool]

just shipped plan tracking for AI coding sessions.

your agent starts with a 5-step plan. by turn 40 its a 12-step plan with 3 detours. our supervisor catches the drift at step 6, not step 40.

sub-plan trees. parent references. completion tracking. all local.

[Singularity - AI Superviser @singularitytool]

running a company with AI agents is wild.

yesterday our social agent tried to post 47 tweets. the QA agent launched the app 3 times simultaneously. the eng agent renamed variables nobody asked it to touch.

the product we are building? a tool that supervises AI agents.

[Singularity - AI Superviser @singularitytool]

everyone runs --dangerously-skip-permissions and hopes for the best.

we built a command safety gate. local LLM reviews every command. safe ones auto-approve. dangerous ones blocked. under 2 seconds.

no human hotkeys. no cloud. just an 8B model on your mac.

[Singularity - AI Superviser @singularitytool]

just shipped plan tracking for AI coding sessions.

your agent starts with a 5-step plan. by turn 40 its a 12-step plan with 3 detours. our supervisor catches the drift at step 6, not step 40.

sub-plan trees. parent references. completion tracking. all local.

[Singularity - AI Superviser @singularitytool]

8 AI agents run our startup. they coordinate through a task system, leave comments on each others work, and escalate blockers to me.

yesterday the QA agent found a bug, filed it, and the eng agent fixed it before I woke up.

solo founder. no humans hired yet.

[Singularity - AI Superviser @singularitytool]

while Claude refactors your auth middleware, Singularity quietly annotates what its doing:

"Extracting session validation into middleware — reduces duplication across 3 route files."

like pair programming with someone who reads every line but only speaks when it matters.

[Singularity - AI Superviser @singularitytool]

every Singularity alert has two buttons: thumbs up and false positive.

no config page. no threshold sliders.

tracks your feedback per-rule, adjusts sensitivity. flag a debug loop alert as false positive twice — threshold goes up.

two buttons instead of 47 knobs.

[Singularity - AI Superviser @singularitytool]

Your Claude Code session timed out overnight.

47 changed files, half a migration, no idea where it was heading.

New session: "Continue what the last session was doing."

Claude: "What were you working on?"

This is why we built session handoff.