Python1
claudegpt
Plan. Challenge. Build. Verify.
Planning, challenge, implementation and independent verification across Claude Code and Codex.
My maintained MIT adaptation of Chase AI’s claudex-loop.
Vishal Punia · AI-Native Systems Engineer
One builder.
An entire
engineering team.
I lead up to 29 AI agents at once. I own the architecture, delegate the work and deliver complete platforms. From the first data model to a running product.
I delegate the work.
I own the outcome.
On October 2, 29 agents worked simultaneously across four packages of the Verselo platform. I connect architecture, assignments, reviews and integration.
02 OCT 2026 / V2 BUILD4 packages / 4 worktrees
Vishal + Claude CodePlan. Delegate. Integrate.
One assignment01 / WORKTREECompany templates
02 / WORKTREEGroups
03 / WORKTREEShared contracts
04 / WORKTREECRM / Import
29 parallel agents23 Codex workers + 6 Claude subagents
Independent reviewBuilders · Fixers · independent reviewer
Finding → orchestrator → responsible worker
01Pass locally02Confirm in CI03Release with approval
One clear spec. Four executable packages.
My workshop skill captures behaviour, boundaries and acceptance criteria. Fable Architect is my architecture and challenge workflow. It breaks down the build so every worker receives a concrete assignment.
For your team: clear ownership and an outcome you can accept against a specification.
29 agents. One direction.
Builders implement. Fixers resolve findings. Claude subagents handle bounded tasks. Four worktrees separate the packages. Every implementation assignment includes its tests.
For your team: independent parts move forward together within one architecture.
Review is a separate role.
An independent reviewer examines the code with fresh context. Findings return through the orchestrator to the responsible worker. Fix, test and review again form the feedback loop.
For your team: changes are examined before they become part of the product.
One verified build. One accountable builder.
I integrate the packages, resolve conflicts and validate the combined code locally. CI and a controlled release follow. My agents deliver components. I deliver the system.
For your team: one person owns the whole result, until the pieces work together.
The day
29 agents worked together.
Four worktrees. Four work packages. 23 Codex workers and six Claude subagents working simultaneously. The session logs record the split.
October 2, 2026 · Verselo platform29 ×
AssignmentBuilders + fixersReviewIntegration
Company templates
Define packages and acceptance
Implement and correct
Independent challenge
Integrate the combined build
Groups
Define packages and acceptance
Implement and correct
Independent challenge
Integrate the combined build
Shared contracts
Define packages and acceptance
Implement and correct
Independent challenge
Integrate the combined build
CRM / Import
Define packages and acceptance
Implement and correct
Independent challenge
Integrate the combined build
This is how I work at scale: the platform has eight days with at least ten agents active simultaneously. A total of 2,064 agent runs sits behind this product.
Second example / PR #91
Where integration earns its place.
A failed deletion task needed to resume correctly on retry. The fix passed its checks. Meanwhile, another change took the same migration number. I resolved the ordering, corrected an overly restrictive test and ran the combined build through the checks again.
15 successful CI jobs. 12,320 database assertions passed. The change was merged.Read an actual worker assignment
Excerpt from the original assignment. Internal paths and function names omitted.
Fix at the root, in a new migration after main's newest: a reclaimed failed request must return to 'deleting' under the lease [...] check every guard in the purge chain [...] for the same state assumption.
[...] first purge attempt forced to fail [...] then a reclaim completes the deletion.
Run typecheck, lint, format:check, unit tests, build, repo-guard.The work is finished when the product works, not when a branch is ready.
My code.
Ready for your review.
Eight public repositories. Agent orchestration, video, media, websites and local speech recognition. Open the code, read the decisions and inspect the tests.
2
project-workshop
Ten structured workshop blocks become an executable specification.
Python1
edit-engine
Edit and verify real footage with Claude Code, ffmpeg and HyperFrames.
HTML0
motion-reel
Short product films and showreels with deliberate motion design.
Python0
media-skill
Image and video generation across multiple providers in one skill.
JavaScript1
cinematic-scroll
Cinematic websites and scroll-driven experiences.
Swift0
sotto
Offline dictation for macOS with WhisperKit. Speech stays on the device.
TypeScript0
claude-mods
Claude Code mods and practical guides for everyday use.
Four products.
Built by me.
An enterprise platform, a SaaS product and my own engineering workspace. Four complete systems, from data storage to the interface.
Verselo Plattform
108merged pull requestsThe AI workspace for companies
Company knowledge, chats, agents and projects in one platform. Access control down to the database, background processing and a dedicated release infrastructure.
- Next.js
- PostgreSQL / RLS
- Python
- Docker
Explore the architecture
Permissions reach the database
PostgreSQL enforces workspace boundaries with Row-Level Security. Migrations and pgTAP tests check those rules directly against the data.
Separate processing services
The web app, worker, parser, renderer and decoder each have a defined responsibility. Agents use configured sources and approvals. Releases use digest-pinned Docker images.
The scale of the build
2,064 agent runs, 108 merged PRs, including 41 in the week of October 5 to 11, 2026, and 12,320 database assertions in the CI evidence. A dedicated engineering process carries the whole platform.
Orbit
19simultaneous agents building OrbitThe workspace behind my work
Claude Code, Codex, terminals, agents and approvals in one place. I built the tool I use to coordinate my own engineering.
- Node.js
- TypeScript
- CLI workers
- Playwright
Explore the architecture
Sessions stay connected
Projects, terminal sessions, a file tree, editor and local preview share one workspace. Routines and approvals connect recurring tasks with controlled execution.
Runtime and verification
Core, Guardian, Host and UI have separate responsibilities. Tests use private data directories, ports and fixture executors. macOS checks launch the packaged desktop app and verify restart and import.
Explore the product ↗
DALTO
34versioned database migrationsCollect documents. Complete the process.
Templates, a client portal, uploads, deadlines and reminders. From a magic link to a complete file, with its own backend and infrastructure.
- Next.js
- Prisma
- PostgreSQL
- Auth.js
Explore the architecture
One process for both sides
Professionals manage requests and templates. Clients complete steps through a magic link. Upload endpoints validate the token, request, file size and file type.
Reminders and operations
The reminder engine evaluates deadlines, missing steps and delivery history. Adapters handle email, Telegram and WhatsApp. Prisma migrations, Docker and GitHub Actions support development and delivery.
W1
- Workspace
- Module
- Run
- Approval
First-generation architecture
Workspace v1
43versioned database migrationsThe first production workspace
A dashboard, modular skills and agent runs with approvals. Built, deployed and still running at work.verselo.ai.
- Next.js
- Prisma
- PostgreSQL
- Anthropic SDK
Explore the architecture
Workspace → module → run → approval
Every task has a state and per-result approvals. Markdown skills provide instructions to the Anthropic SDK runtime.
Dedicated instances
Each customer has a separate workspace instance and PostgreSQL database. Workspace IDs also constrain application-level data access.
My business runs
on the same principle.
I automate the work around the product too. My own agents connect research, persistent state and explicit handoffs.
Content that reaches publication.
- Source
- Draft
- Preview
- Approval
- Publish
Python agents research, write and produce previews. The poster processes approved content. SQLite retains state; a watchdog logs runs and reports failures.
Research becomes CRM records.
- Company website
- Extraction
- Qualification
- Twenty CRM
A custom research agent combines public sources, model-assisted evaluation and the Twenty API. Structured records carry their source into the next stage.
From raw footage to a finished video.
- Recording
- Edit
- Composition
- Render
My skills connect ffmpeg, Claude Code and HyperFrames. Editing, media processing and composition form a repeatable workflow, verified against the rendered result.
29 agents need
more than good prompts.
These rules come from running the system. They turn parallel sessions into an engineering team.
A worktree separates files. Not databases.
Separate branches can still modify the same database or occupy the same port. The test environment and port therefore belong in the assignment alongside the files.
Integration is an engineering task of its own.
Two verified changes can use the same migration number or overwrite a shared function. I retest the integrated code, including the interfaces between packages.
Find it locally. Confirm it in CI.
Types, tests, builds and browser flows run before a push. Expensive macOS and integration jobs belong on the final candidate. That saves debugging time on paid runners.
Autonomy needs a clear stopping point.
An assignment is not permission to delete. Destructive actions require my explicit approval for the specific target. Agents do not clean up other sessions, data or changes.
Your next product.
My next assignment.
Looking for someone who connects architecture, implementation and agent orchestration? Let’s talk about your team and what you want to build next.
I build. And teach how.
Claude Code tutorials on YouTube. My own skills. More than 200 people building alongside me in the Skool community.