ARCHITECTURE & DEVELOPERS

Build on DRENTIR.

Every great operating system became great because developers built on it. DRENTIR is designed as a platform from day one: tooling for specialized agents, connectors to the software you already use, and APIs for programmatic control. If you build software, there will be a way in.

Developer APIs and SDKs are not public yet. This page documents the architecture direction and technical preview.

BUILDING BLOCKS

Three ways to build.

Different builders need different doors into the system.

01

Agents as primitives

Define specialized agents in code — their role, their tools, their data scope, their permission boundaries. Version them, review them, roll them back like code.

02

Connectors

Plug into the software your users already run: email, calendar, CRM, documents, messaging. One connector makes every agent smarter.

03

Open schemas

Plans, tasks, results and the audit trail share one documented schema. Anything you build speaks the same language as everything we build.

SYSTEM PLANES

Five planes, one system.

DRENTIR is organized as planes — each with a clear job, each crossing the others only through defined boundaries.

01

Interface plane

Where you meet DRENTIR: conversation, voice and the Command Center. Everything you see and say starts here.

02

Agent plane

Where work gets planned and coordinated: specialized agents, plans, orchestration. The reasoning layer of the system.

03

Tool plane

Where DRENTIR touches the world: connectors to email, calendar, CRM, documents, messaging, browser and files.

04

Model plane

Where thinking happens: local models for privacy and speed, cloud models for capability — routed per subtask.

05

Control plane

What keeps it yours: permissions, approvals, policies and the audit trail — crossing every other plane.

ARCHITECTURE

How a request travels.

One request crosses the planes top to bottom; the control plane watches every crossing.

You

A goal in plain language — spoken or typed.

Interface plane

Understands the request, manages the conversation.

Agent plane

Builds the plan, coordinates specialized agents.

Tool plane

Calls connectors: email, calendar, CRM, files, browser.

Model plane

Routes each subtask to the right model — local or cloud.

Control plane — spans every layer

Permissions, approvals, policies and the audit trail apply at every step.

Architecture direction — the system is a working prototype, internals are evolving

EXECUTION CONTEXT

Everything an agent may do — declared up front.

Three sketches of the design direction: who the agent works for, what it is built from, and what it is never allowed to do.

review.context
context "weekly-review" {  owner: maya@studio  workspace: studio.ops   permission: PREPARE   # draft, never send  data_scope: [mail.read, calendar.read]   escalate {    before: [send, spend]    on_urgent: notify.owner  }}
ExecutionContext — whose work it is, what it may touch, where it must ask.
triage.manifest
manifest "triage-agent" {  version: 0.3.1  runtime: drentir.agent/1   tools: [mail.read, crm.lookup]  models: {    default: local.qwen3,    fallback: cloud.claude  }   limits {    max_cost_per_day: 2.00 USD    max_actions_per_run: 25  }}
Manifest — what the agent is built from: tools, models, limits.
outbound.policy
policy "outbound-comms" {  deny: [send.email, send.sms]  allow: [draft.email, draft.sms]   require_approval:    [spend.money, change.records]   audit: everything   # no silent actions}
Policy — what is never allowed without a human.
Illustrative sketches of the design direction — not final syntax

DIRECTION

Agents as code, permissions as code.

The shape we are designing toward: an agent is a small, reviewable definition — what it does, what it may touch, and where it must ask.

triage.agent
agent "triage" {  on: inbox.new_message   permission: PREPARE          # draft only, never send  tools: [mail.read, crm.lookup]   escalate {    before: send              # always ask first    on_urgent: notify.owner  }}
Design direction, subject to change — illustrative sketch, not a final syntax.

PRINCIPLES

Built the platform way.

Permissions from line one

Anything you build inherits the permission model — your agents can never outgrow the boundaries they are given.

Observable by default

Your agents write to the same audit trail as everything else. Debugging stops being guesswork.

Humans stay in the loop

Escalation is a primitive, not an afterthought: define where your agent must ask, and it will.

ROADMAP

What builders get first.

The platform opens up in stages. Early access members shape each one.

Agent definitions

Define and deploy agents through the sketch above. This is the design direction under active development right now.

Read API first

Programmatic access to plans, tasks, results and the audit trail — before anything can write.

Connector SDK

Build connectors others can install: email, calendar, CRM, documents, and beyond.

Write API and sandbox

Mutations and a safe test environment arrive after the read surface is solid.

No developer API shipped yet

QUESTIONS

Asked by builders.

Want to build on DRENTIR?

Join early access and tell us you build — developer previews go to members first.

Join Early Access

DRENTIR Website Guide Ask me anything about DRENTIR. Not connected to your personal AI environment.