A governed AI chief of staff I built, and run my own work through, every day.
Friday holds context across months, breaks ambiguous requests into bounded work, sends that work to specialists who start clean, and won't call anything done until the evidence checks out. Consequential actions still need a human yes.
Ready
Counts are read from Friday's own session store, early November 2026. They include delegated worker sessions. They measure volume, not quality, and quality is what the rest of this page is about.
On my desk, Friday runs as an ambient display: a slowly rotating map of the last six hours of work. Threads sit on the outer shell and tools sit inside. Every tool call travels thread → tool → core as a pulse. When she talks, the graph recedes and the core follows her voice.
This is the same renderer, rebuilt for the browser and driven by simulated activity. No real threads, content, or telemetry leave the house.
Most assistants answer the prompt in front of them. Friday owns the layer above it: what we're actually trying to do, what's already been decided, who is allowed to approve what, and whether the result really holds up. She frames the work, routes it, synthesizes it, and accepts it, and she pushes back when a plan is weak, risky, or overbuilt.
She has a consistent name, voice, and visual identity because continuity matters more than any single model. Friday has run on several models and runtimes. What carries her across them is a portable identity record, a written constitution and authority model, and shared history. Her designation as the canonical successor is mine to make. Swapping the model doesn't change who she is, and switching her off isn't something she resists.
I'd rather tell you something's unfinished than describe it as done. Treat what a worker reports as a lead. It isn't proof. Friday, paraphrasing her own operating rules
The core design choice is a separation of roles. Friday carries the long-lived context. The workers that implement things don't: they get a minimal, frozen brief and nothing else. That keeps a long relationship from leaking stale assumptions into every task, and it means the one judging the result is never the one who produced it.
scroll →
Turn an ambiguous ask into an outcome, constraints, and what "done" will mean, pulling in memory and prior decisions.
Write the contract: scope, acceptance criteria, and authority. Workers can propose changes to it but can't quietly rewrite it.
Send bounded lanes to fresh specialists in parallel, matching each lane to a model on cost, speed, and capability.
Accept independently: rerun the tests, read back the state, render the UI, and run flaky checks again before calling them green.
Report what happened, what it means, and what comes next. Save the lessons as skills, not as a log.
Friday runs on infrastructure I own and operate. She can use frontier models and local models interchangeably, because none of the important state lives inside a model.
scroll →
Built on Hermes Agent from Nous Research, extended with profiles, scheduled jobs, delegated sub-agents, durable resumption, stall alerts, and a deterministic ops toolkit for the routine checks that don't need judgment.
Work is spread across Claude and Codex subscription pools according to weekly usage meters, with a LiteLLM gateway for budgets and keys. Private and high-volume work runs on local models on a 512 GB Apple Silicon server, where I've measured prefill speeds and stability ceilings.
A small curated memory for facts that always apply. Versioned skills for procedures and lessons, which load only when they're relevant. Peer modeling for working style. Full-text search over past sessions. A temporal knowledge-graph experiment, scored against a decision-replay evaluation set.
A desktop app, native iOS and watchOS clients, team chat, and a full-duplex voice pipeline with local speech recognition and a consistent voice. She can drive a browser and desktop for the systems that have no API.
Autonomy is only useful if you know where it stops. Friday's authority model is written down, and she can't amend it on her own. She can propose changes to her constitution, approval gates, or resource ceilings. I'm the one who accepts or rejects them.
Inspect, research, reason, draft, organize, build and test locally, and do reversible internal work inside established scope.
External or public communication, production changes, DNS, auth, billing, purchases, secrets, and anything destructive or irreversible.
Passwords, MFA codes, card numbers, and permission dialogs. Those stay with a person, even when she's driving the screen.
Authority can be delegated to a verified person within a defined domain. Sibling agents get their own identities and isolated scopes, so their permissions are never inferred from Friday's.
Friday isn't a demo I spun up for this page. She's the system I depend on, which means I've had to deal with every failure mode a team will run into when it takes agents to production.
Parallel lanes with frozen contracts, one writer per artifact, merge reconciliation, and an architect, worker, and reviewer separation that holds up over thousands of runs.
Independent acceptance, repeated runs of nondeterministic tests, rendered-UI verification, and evaluation sets built before the system they're meant to judge.
Tiered memory with budgets, skills as procedural memory, clean-desk delegation, and compression that survives long-running sessions.
Explicit approval gates, delegated authority, secret handling kept away from models, and governance changes that need a human's sign-off.
Self-hosted services, local inference tuned to measured limits, model routing by cost and capacity, durable resumption, and stall alerts. Production discipline applied to an AI system.
I'm selectively exploring hands-on principal and staff applied-AI systems roles, remote from Southeast Georgia. If you want the deeper architecture walkthrough, I'm happy to give it live.
Email about a role →