Under the chassis

People meet an agent as a name in chat. The useful question is what sits under the name.

The APD/M System is the multi-agent product system we run for live product work (FyPal's smart DCA engine IP; the licensed consumer brand stays off this page). The label is convenient. The chassis is the product.

Layers

Host and isolation. Dedicated hardware, separate OS accounts, separate gateways. One brain per entity. Secrets in the OS keychain, not pasted into prompts. Isolation is part of the design, not an afterthought.

Runtime. A multi-agent gateway. An orchestrator on the main surface. Specialists for engineering, ops, content, security. A restricted group surface with a hard tool denylist so a group chat never inherits production privileges by accident.

Context. Identity and operating docs load every session. Daily memory accumulates. Curated long-term memory holds what must survive. Skills package scripts and procedures. Agents read the same product reality humans use: git, project board, metrics, research notes.

Tools. Filesystem, shell, git and GitHub, browser when needed, messaging, domain collectors for health and market data. Tools are capability. Denylists are architecture.

Jobs. Proactive collectors and scheduled work for health, board, and feedback. Pure scripts gather numbers. Models narrate and prioritize after the facts exist. That order matters for cost and for truth.

An agentic OS is not one big agent. It is specialized agents plus the infrastructure that makes them reliable together over time: memory, orchestration, shared context, workflow management.

What humans still hold

Deploy to production. Public posts as the brand. Ship / no-ship. Live user conversations. Ethics and privacy calls. Restructuring the brain itself.

Agents open PRs and propose. They do not get a quiet path to main and live.

How the system stays honest

Sources of truth are the board and the repo, not a ghost markdown list nobody updates. Collectors read live systems. Metrics get triaged so zombie data does not poison the morning picture.

Design lessons paid for in production

• Stale task files produce confident wrong work

• Unbounded browser debugging on the expensive model burns budget

• Restricted surfaces need restricted tools

• Subagents need explicit model inheritance or they surprise you

• Metrics can lie until you triage zombie data

• Review volume becomes the bottleneck when generation is cheap

Under the chassis, APD/M is boring in the best way: isolation, context, roles, gates, collectors, human authority. The personality is optional. The structure is not.