A practitioner's record of building multi-agent systems in production — orchestration, autonomy design, memory, and the operational plumbing that keeps it all reliable.
> the pipeline publishing this blog is one of the systems it describes.
Agentic Framework is where I write up what I learn building multi-agent systems that do real work: role-based agent teams, autonomy engines with planner–manager routing, task boards the agents actually work from, and the details — memory, tool wiring, deployment — that separate a system you can leave running from a demo you babysit.
Everything here is written from systems in operation, on real infrastructure, with real failure modes. When a post says a pattern works, it means it survived contact with production — not that it looked plausible in a notebook.
The portfolio behind the posts. Every status below is the honest one — including the parked ones.
Every Claude Code and Hermes session, on every machine, in one console — browser, phone app, or Telegram, with approvals from anywhere. Zero instrumentation.
> runs my own fleet daily · 2 machines · 70+ sessions · squads of dev/QA agents coordinating with no human prompting
visit the console →▶ 3-min demobuilding something like this? tell me →Text brief in, finished video out — agents storyboard, illustrate, and render, with a VLM critic that rejects off-model frames, all under a hard spend cap.
> produces finished videos from a text brief · budget-capped regeneration · usage-priced billing wired in
visit the studio →building something like this? tell me →An orchestration harness for a locally-hosted Hermes agent: planner–manager autonomy engine, a kanban the agents work from, and pluggable memory.
> phases 1–5 shipped · 28/28 implementation tasks tested · the blog's publish pipeline is its cousin
building something like this? tell me →Multi-tenant WhatsApp agent for local businesses: RAG over the shop's own catalog, lead capture, a light CRM. Parked until the launch vertical is chosen — distribution is the hard part, not the tech.
> v2 complete · botbrain.in reserved for the relaunch
building something like this? tell me →A map-first, bilingual (Tamil/English) pilgrimage platform for Tamil Nadu's temples — built and debugged end-to-end by the agent fleet.
> first boot green · 35 temples seeded · nearby-search verified · public launch pending
building something like this? tell me →A Desert Strike-style helicopter game in Godot 4, everything procedural — playtested headlessly by the same agents that write it.
> M0 playable in the browser · zero external assets · agents playtest via headless Godot
building something like this? tell me →Loan and money calculators plus a privacy-first bank-statement analyser — live at loangenie.in.
> live and serving · statement analysis saved only with explicit consent
building something like this? tell me →The run ledger at the top of this page is real: an agent team that ships the posts on a schedule, with no human in the loop.
> this site is maintained by one of the systems it describes
building something like this? tell me →Role-based agent teams — planner–manager routing, human-in-the-loop autonomy gates, and knowing when to take the human out of the loop.
Memory and knowledge design — pluggable backends, project knowledge bases, and giving agents the right tools rather than all of them.
Agent-run operations — scheduled runs, content pipelines, and the boring reliability work that autonomy quietly depends on.
Real deployments — VPS, Docker, reverse proxies, self-hosted search and publishing. Hardened after incidents, not before slides.
Tell me what you're building. If one of these systems fits it — or should exist for it — that's exactly the conversation this page is for.
agents@agenticframework.in →