People set direction
The operator chooses goals, taste, risk tolerance, and review points. The system is designed to make that direction visible instead of burying it in chat history.
Public manual
Night Shift, explained like a field manual.
Night Shift is an agent-company operating layer: humans set direction, scoped agents do useful work, and shared source-of-truth rails keep the work inspectable. This public manual explains the model without exposing live private state.
Agents do scoped work
Workers take bounded assignments, use the tools they are allowed to use, and leave evidence behind. They are treated as accountable contributors, not mascots.
Truth lives in rails
Kanban records execution, the wiki records durable project knowledge, and the Control Room reads from safe snapshots rather than inventing state.
Operating model
How the pieces fit
The simplest path is: human direction flows to an orchestrator, the orchestrator routes scoped work to capability agents, and the results are recorded in durable rails. The Control Room reads and explains; it does not become a hidden command bus.
- Human direction A clear brief: what matters, what is risky, and when a person must review.
- Orchestrator Breaks work into smaller assignments and routes them to the right capability.
- Capability agents Implement, inspect, research, review, or document one scoped piece of work.
- Kanban ledger Tracks ownership, dependencies, blockers, evidence, and handoffs.
- Wiki memory Keeps decisions, specs, gotchas, and runbooks available after a session ends.
- Control Room Presents a read-only command-room view of selected public or redacted information.
Night Shift
An operating layer for agent-assisted work.
Night Shift is the pattern around the agents: a way to keep useful work moving across products, research, maintenance, and documentation after the human operator has set direction. It is not a single bot or a magic autopilot. It is a company-like operating model where humans provide intent, agents carry out bounded pieces of work, and reviewers decide when risk is acceptable.
Why it matters: the system is easier to trust when it looks like accountable operations instead of improvisation.
Hermes
The runtime that gives agents shape.
Hermes is the operator surface for agent profiles, tools, skills, scheduled jobs, gateways, and worker sessions. In this manual we describe those concepts at a public level: profiles separate responsibilities, tools define what a worker can touch, skills keep repeatable procedures close at hand, and scheduled jobs let recurring checks run without a person retyping instructions.
Why it matters: capability boundaries make an agent workforce easier to review, recover, and improve.
Kanban
The execution ledger, not a vibes board.
Kanban is where work becomes inspectable. A task can be decomposed, assigned, blocked for a human decision, handed off with metadata, or completed with evidence. The important public idea is accountability: workers should not merely claim progress; they should leave enough context for the next person or reviewer to understand what changed.
Why it matters: the ledger turns autonomous work into a chain of reviewable decisions instead of a fog of chat messages.
Wiki
Durable memory for projects and runbooks.
The wiki is where long-lived project truth belongs: build pages, specs, design decisions, gotchas, status notes, and runbooks. It is not a dump of every conversation. It is the cleaned-up knowledge layer that helps future workers start from the same map.
Why it matters: agents need stable context that survives model sessions and keeps project intent from drifting.
Control Room
A public shell around a private cockpit.
The Control Room is a presentation layer. Public pages explain the operating model and selected sanitized examples. Private cockpit panels, when enabled, should read from approved snapshots with freshness, source, error, and confidence states. The Control Room is not a second task system and does not replace the wiki or ledger.
Why it matters: readers get a useful mental model without exposing live operations.
How to read the manual
A guide to the system, not a window into the control plane.
Start with the mental model
Read the manual as a map of responsibilities. Humans own intent and approval. Hermes supplies the runtime. Kanban records the work. The wiki preserves durable context. The Control Room presents safe summaries.
Follow the evidence trail
Every serious piece of agent work should leave a trail: what was requested, who worked it, what changed, how it was checked, and what still needs a human decision. Public examples show the shape without exposing real operations.
Separate public education from private operation
The manual can explain concepts, roles, and review patterns. It should not publish live queues, private logs, private routes, filesystem locations, credential classes, or controls that could operate a real system.
Kanban lifecycle
How agent work moves from request to accepted handoff.
A public reader does not need internal queues to understand the discipline. The important pattern is that each step narrows ambiguity and leaves enough evidence for the next reviewer.
- Intake A goal is stated with the outcome, risk, and review expectations.
- Decompose Broad work is split into small assignments that can be owned and checked.
- Assign A capability receives only the scope and tools needed for that assignment.
- Work The worker changes code, writes docs, researches, reviews, or prepares a handoff.
- Evidence Commands, assumptions, changed artifacts, and verification results are recorded.
- Block or review Human judgment is requested when the next step changes risk or trust.
- Complete Accepted work closes with a concise handoff so the next reader can pick up context.
What belongs where
One system, different kinds of truth.
Chat Fast conversation, clarification, taste checks, and short-lived collaboration.
Kanban Assigned work, dependencies, blocker questions, review gates, and evidence.
Wiki Durable decisions, specs, runbooks, gotchas, and project memory.
Control Room Read-only summaries, public explainers, and eventually redacted operational snapshots.
Fictional task card
What evidence can look like.
TASK-123 ยท review required
Draft a public-safe project explainer
- Owner
- agent:writer
- Evidence
- Changed copy, checked build, listed assumptions.
- Blocker
- Human approval needed before publishing.
The shape is real; the details are not. Public examples use aliases so the manual teaches the workflow without leaking live work.
Capability roster
Capability-first, not celebrity bots.
Orchestrator
Turns a broad goal into small, reviewable assignments.
Should not quietly perform risky implementation work just to keep things moving.Implementation worker
Changes code, docs, fixtures, or tests inside a defined scope.
Needs review for high-impact or public-facing changes.Reviewer / Sentinel
Checks safety, quality, privacy, and readiness before work is treated as accepted.
Does not rubber-stamp unclear evidence.Scout / Researcher
Collects facts, options, sources, and trade-offs.
Should separate sourced claims from assumptions.Project specialist
Carries context for a product area, workflow, or domain.
Should still use shared rails rather than private memory alone.Public / private boundary
Useful explanations without operational leakage.
Concepts, sanitized workflows, fictional examples, high-level role descriptions, and design principles.
Live operational summaries, approved redacted snapshots, freshness indicators, source status, and reviewer-only context.
Credential values, logs, raw snapshots, private paths, live task bodies, private contact identifiers, admin controls, or deployment operations.
Sources of truth
The manual explains the map; the rails keep the truth.
Night Shift works because each kind of information has a home. Kanban carries execution state, the wiki carries durable knowledge, repositories carry code, and private snapshots can summarize selected state without becoming the authority.
The Control Room should point back to the real system of record instead of storing a second copy of project truth.
Public pages can explain the shape of a workflow, but live counts, exact blockers, raw handoffs, and operational details stay private.
When private snapshots exist, each panel should say when it was collected, where it came from, and whether the data is current or partial.
If a collector cannot read a source, the honest public-safe state is unavailable, not a confident guess.
Sanitized examples
How agent work shows up across projects.
The examples below describe safe patterns, not live project status. They show how the operating model can support builds, research, and maintenance while keeping private work private.
Product build
Agents can draft specs, inspect implementation details, write tests, and prepare review notes while a human keeps product taste and launch risk in view.
Publishing, pricing, customer promises, and production changes remain human-reviewed.Research track
A scout can gather sources, compare options, and turn a messy question into a sourced brief that later workers can reuse.
Claims, recommendations, and irreversible decisions need a person to check the evidence.Maintenance loop
Scheduled checks can surface stale docs, dirty repos, failed jobs, or missing evidence without silently changing the system.
Cleanup, deploys, account changes, and destructive fixes require explicit approval.Reviewer gates
Where autonomy intentionally stops.
Secrets and access Anything involving credentials, sessions, private contacts, or access boundaries stops for review.
Production and infrastructure Deploys, process restarts, DNS, databases, payment systems, and live platform mutations are not public-docs actions.
Public claims Commercial, legal, medical, financial, affiliate, analytics, or launch claims need a human source-of-truth check.
Destructive work Deletes, resets, force pushes, cleanups, and irreversible state changes need a separate approved path.
Glossary
Short names for the operating model.
Profile A scoped Hermes identity with a role, toolset, and operating boundary.
Skill Reusable procedural context that teaches a worker how to do a recurring kind of task.
Snapshot A point-in-time read model derived from private sources, with freshness and error state.
Handoff The concise evidence package a worker leaves for a reviewer or downstream worker.
Safety and trust
Useful public docs, private operations.
The public manual should make Night Shift understandable without publishing the machinery that operates it. The safe substitute for live data is a clear concept, a sanitized pattern, or a fictional example.
- Public pages use fictional identifiers such as TASK-123 and generic agent names.
- Credential values, logs, raw runtime snapshots, private paths, contact identifiers, and live task bodies are never published.
- Mutation controls belong behind explicit approvals, not in the public manual.
- Unknown or unavailable state is shown honestly instead of inferred from stale data.