Harness blueprints
A coding agent, a personal assistant and a support desk look nothing alike from the outside. Inside, they are the same parts: a loop, tools, context, state, a workspace, a surface and sometimes a team. What changes is which parts carry the weight. Each blueprint below lists the primitives a harness uses, what you write yourself, and how it deploys.
If the parts are new to you, read First principles first. If you are about to build something and want to know whether it already exists, check the primitives map.
Same parts, different weight
Read a column top to bottom and the pattern shows: every harness has a surface and context, most have state, and only some need a workspace or a team. The system column is where the agent stops being one conversation and becomes a service.
| Harness | Surface | Context | State | Workspace | Team | System |
|---|---|---|---|---|---|---|
| Coding agent | Terminal or web client over a socket; plan approval with pushAndWait | Repository guides as skills, the task list, compaction | Transcript, permission decisions | A sandbox, or the host directory behind permission gates | Read-only explorer and reviewer subagents | Optional: Foundry for background runs |
| Personal agent | Telegram, email or WebChat | Pinned preferences and live state at the tail | Memory of people, events and notes | A browser, documents, a sandbox | A research subagent | One instance per person, schedules, sleep, approvals |
| Support desk | The ticket thread or chat widget | The customer's record as live context | Forms, facts and goals per conversation | Usually none: MCP or tables over your systems | A read-only research subagent | Instance per customer, transmissions, playbooks, subscriptions, classifiers, approvals |
| Voice agent | Speech: a pipeline or a realtime model | Runtime context, refreshed silently | Forms that work over speech | None in the voice loop | A fast front agent and a heavy worker over mesh | Durable work delegated to Foundry |
| Research & documents | Chat, with files handed over by present | Skills in /skills; the script library | VFS snapshots between runs | Working environment, a REPL, format adapters | A reviewer subagent | VFS persistence, background work |
| Multi-agent workforce | The inspector, plus your own briefing UI | One prompt per role | Workspace entries, tasks, the shared inbox | Shared VFS paths passed by reference | Spawn, mesh, subagents | Subscriptions that provision agents, schedules |
The blueprints
File and shell tools behind permission gates, a plan the person approves, a task list, repository skills, subagents, and a sandbox when the code is not yours to trust.
glove-coreglove-executionState · Time · ArrivalPersonal agentAn assistant that works while you don't: a chat channel, memory, reminders and sleep, a browser and documents, email, approvals before anything irreversible.
glove-foundryglove-memoryglove-executionIdentity · Arrival · OversightSupport deskAn agent per customer that wakes on tickets, triages with a classifier before any model runs, collects details with forms and waits for a human before refunds.
glove-foundryglove-classifierglove-memorySurface · TeamVoice agentA fast front that keeps talking while a worker does the tool work, on a speech pipeline or a realtime model, with durable work handed to Foundry.
glove-voiceglove-voice-s2sglove-meshWorkspaceResearch & documentsAn analyst that computes in code, writes PDFs, decks and spreadsheets as files, and looks at its own output before handing it over.
glove-working-environmentglove-env-documentsglove-jsTeam · Identity · TimeMulti-agent workforceAgents provisioned by events, coordinating through workspace entries, tasks and mesh, with one inspector over all of it.
glove-foundryglove-meshHow to read a blueprint
Each page opens with a bill of materials: for each of the three layers, the parts the harness uses and the primitives it takes from each, then the parts it doesn't need. The dashed box is what you write: usually your domain tools, your prompts, and the adapter for your own channel or data. Everything else is already built.
Then comes one request or event walked through the harness, a skeleton you can start from, the things people most often hand-build for that shape (and what to use instead), and the steps to deploy it on Kuiper.
To build the same parts up one at a time instead, follow Build it up, which takes one agent from a single tool to a deployed Foundry system in eight steps.