Most of what I build these days ships inside companies, under NDA and their IP. No repos, no screenshots, no client names. What I can tell you is what each agent does, how it is built, and whether it is live. Ask Neo-Oruga (red pill) if you want the longer story.
Agents built for clients
Persistent memory per person, so it remembers what you talked about weeks ago. A personal assistant that summarizes your inbox, briefs you before meetings and keeps reminders. Deep search across the company’s Google Workspace (Drive, Gmail, Calendar) with per-user isolation, so everyone only sees what they already have access to. Writes only behind approval gates.
Tech: Claude Code harness, Vertex AI, Slack, Google Workspace APIs, GCP
Status: In production for a client, under NDA.
A control plane that decides who may do what, a data plane that executes, and a web UI. Trusts everyone only up to their clearance and fails closed. Other agents plug in as vassals: they keep their own code and runtime, it becomes the one entry point and the one place where access is decided.
Tech: TypeScript, Claude on Vertex AI, GCP, capability tokens
Status: Code complete, deployment gated. Under NDA.
When someone joins, the agent provisions the accounts, writes them a personalized welcome, walks them through role-specific onboarding in Slack and follows up until each step is done. Covers IT, pipeline and business operations, and the same rails handle role changes and offboarding.
Tech: Python, Claude on Vertex AI, Google Workspace admin APIs, Slack, GCP
Status: In production for a client, under NDA.
Audits the state of access and policy (who has what, which rules are stale), never the content people produce. Report-only by design: findings land as a Slack digest and an HTML dashboard, with any enforcement deferred behind dry runs, a double key and an undo. Reads the onboarding agent’s memory through a one-way wall.
Tech: Claude Code harness, Python, Slack, Google Drive (create-only)
Status: Designed and documented (PRD + solution design). Under NDA.
A small internal platform where teams file requirements and authorizations flow to the right approver, with dashboards and an action assistant that drafts and explains but never writes on its own. Deliberately a platform, not an autonomous agent.
Tech: Next.js, TypeScript, SQL, Claude on Vertex AI, Cloud Run behind identity-aware proxy
Status: In production for a client, under NDA.
Reads the channels and the related inbox for a production, extracts decisions, risks, blockers and follow-ups, and rebuilds a dashboard on every run with only the delta since the last one. Leadership gets the signal without reading a thousand messages.
Tech: Python, Claude, Slack API, Gmail API, Google Drive, static dashboard
Status: In production for a client, under NDA.
Pulls recent deal and contact activity from the CRM and posts a clean feed into the sales channel, with an on-demand scan command for when someone needs the latest before a call.
Tech: Python, HubSpot API, Slack
Status: In production for a client, under NDA.
A Gmail watch fires a Pub/Sub event, the agent maps the email into a structured row in the team’s tracking base with a small fast model, labels the thread, and never replies, forwards or deletes anything. Built as a skill-per-feature harness so the dev team can add the next feature the same way.
Tech: Claude Haiku via claude -p, Gmail API, Cloud Pub/Sub, AirTable, systemd on a GCP VM
Status: In production for a client, under NDA.
A CLI plus dashboard that compares every project’s folders and assets to a policy file, proposes fixes in two lanes (cosmetic labels always safe, path-sensitive slugs guarded), keeps an undo trail, and reports how many releases behind each pipeline addon is with the missing release notes inline.
Tech: Python, Click, official pipeline SDK, GitHub API, Cloud Run
Status: In production for a client, under NDA.
My own harness, a.k.a. the modifications to Claude
Everything above is built by me plus a heavily customized Claude Code. These are the modifications that make it work like a team instead of a chatbot.
Layer one is a Karpathy-style LLM wiki: curated, interlinked markdown over immutable sources, with ingest, query and lint. Layer two is verbatim recall through a local memory palace that mines the project and past sessions with zero API calls. An awakening protocol runs at every session start and a compact hook files what was learned before context is summarized. Personal, work and harness-only modes.
Tech: Claude Code hooks and skills, MemPalace, markdown
Status: Private. Installed in every project, never published.
A pre-edit hook blocks any change to a code file until the agent declares what it is about to change and why; a post-edit hook records what actually changed; a stop hook reports every entry left without a reason or a verification. One JSON per file, mirror tree, local only, never committed. Applies to the main agent and to every subagent.
Tech: Node, Claude Code PreToolUse / PostToolUse / Stop hooks
Status: Private. Running on every project on this machine.
A sealed subprocess runs an isolated agent against a different model provider with its own empty config, so nothing from the main harness leaks and the main session stays on Anthropic. Claude writes the technical spec, validates the output and operates the generators; the tunnel writes image and video prompts and design-forward frontends. Every run reports real cost.
Tech: Node, claude -p, isolated CLAUDE_CONFIG_DIR, Kimi K3
Status: Private. Used for the covers on this very page.
Private · patterns public
Autonomous build loops and adversarial review
Goal, boundaries, verifier, evolve cadence. Then let it run.
A loop contract per project (state, logs, backlog, gates) that an agent reads top to bottom, takes the next unblocked item, ships it with tests, and updates its own state before ending the run. Paired with an adversarial review skill that spawns one refuter per claim, each trying to kill it with reproducible evidence, and only lets through what survives. The Orugas game on this page came out of 43 runs of this loop.
Tech: Claude Code skills, Agent Teams, Workflow scripts, git worktrees
Status: Private, patterns published in claude-skills.
Multi-project ticket board with epics, acceptance criteria and per-project keys, where each desk is a Claude Code skill running as a real headless agent. A scheduled drainer works the queue, a planner must write acceptance criteria before anything ships, and a big red emergency stop halts the office from any page.
Tech: Next.js, Postgres, Drizzle, claude -p, Windows Task Scheduler, Vercel
Status: Private repo, live for personal use.
When a turn ends, a one-line spoken summary in first person tells me whether something got done or the agent is waiting, only in the session that asked for it and only if the webcam confirms I am at the desk. Plus a general brain that remembers across every project, and a mesh where each laptop’s Claude Code is an agent that wakes up when mentioned in a channel.
Tech: ElevenLabs, Claude Haiku, Claude Code UserPromptSubmit and Stop hooks, Buzz relay
Status: Private. Running daily.