Guides

One file, one agent; the access boundary; execution scopes; portability

Everything an agent can do and what you can do with it, then the mechanisms underneath.

Walkthrough of creating an agent and every per-agent setting: identity, model, context, tools, messaging, limits, serving

CLI: Agent settings

The .adf files new agents are made from, the three Studio ships, what an instance carries, and how to edit, share and pick a template

CLI: Templates and skills

The agent state model and how to control it: states, transitions, loop modes, turn tools, and the birth-to-death lifecycle

The virtual filesystem: README.md, mind.md, soul.md, protection levels, templating, meta keys, local tables, adf-file:// links

CLI: Files

Every built-in tool plus the access-control model (enabled/visible/restricted), locking, and cross-cutting parameters

Skill package conventions: first-party catalog, directory layout, SKILL.md frontmatter, the runtime indexer, security boundary

CLI: Templates and skills

The sandbox all agent code runs in: contexts, security model, globals, packages, transforms, timeouts, state persistence

The global adf RPC proxy in code execution: calling convention and every special method (model_invoke, identity, attestations, events)

File-level trust boundary for privileged code: restricting methods and tools, authorization flow, governance patterns

Where agents run commands: shared/isolated/external containers and host access, approval policies, security posture

The Computer tab: each isolated agent's visible desktop with a managed Chromium session, Openbox/tint2 window management, computer-use CLI tools, user takeover for security checks, and portable browser profiles

How an agent drives its own desktop: look at the screen, click, type, open apps and files, share the screen with its principal, using compute_exec, fs_transfer, xdotool, scrot and xclip

Connecting external MCP tool servers: installation, credentials, OAuth, per-agent attachment, and tool lifecycle

CLI: MCP servers

Event-driven agent activation: trigger types, targets, scopes, filters, timing modifiers, and lambda event payloads

Scheduling future events: one-time, interval, and cron timers, timer lambdas, lifecycle, and missed-timer handling

CLI: Loops and timers

Deferred and asynchronous tool executions in adf_tasks: creation paths, statuses, querying, and resolution

Inner loops (side loops): up to 16 named loops inside one agent, each with its own transcript, goal, tool subset and pacing

CLI: Loops and timers

Transform a normal conversational provider request with a workspace lambda

Managing the loop (conversation history) and mind (working memory): compaction, audit, context blocks, long-running strategies

CLI: Context

The mind wiki pattern: OKF frontmatter format for mind pages, index/log contract, citation scheme, audit retrieval recipe

The full messaging system: DID mesh messaging, delivery, security, and per-platform channel adapter setup walkthroughs

Channel adapter contract: addressing, content modes, form render contracts, credentials and self-setup, offline catch-up, group context, chat_info

CLI: Channels

Agent-level contact management patterns: what to remember, whom to trust, how to route; runtime primitives vs agent policy

Finding agents across machines via mDNS: announcements, discovery merging, and LAN troubleshooting

Persistent bidirectional agent connections and NAT traversal: hot/cold paths, auth, reconnection, binary frames, backpressure

HTTP serving through the mesh server: public files, shared workspace globs, and lambda-backed API routes

Custom middleware lambdas in the message and request pipelines: inbox, outbox, fetch, and route interception points

Every outbound connection Studio and the daemon make, who starts it, and how to turn it off.

The runtime security model: trust boundaries, defense-in-depth layers, and hardening controls

The identity and secret-protection stack: owner/runtime/agent DIDs, envelope encryption, sharing, claiming, and attestations

CLI: Identity and security

adf_logs reference: entry schema, what gets logged, filter/retention config, ring buffer, on_logs trigger, querying

The real-time event stream of an agent's own activity: taps, loop protection, replay, and the durable-tap recipe

Reference for every umbilical event type: envelope, per-namespace payloads, and stability tiers

Task-oriented Knowledge Base for humans and agents using ADF capabilities

Run visible Linux desktop applications in isolated compute, transfer files, and validate visual results without assuming generic desktop automation

General software-engineering practice for ADF agents and humans: isolated workspaces, GitHub access checks, identity-owned credentials, bounded delegation, and accountable delivery

The ADF file format and the ALF message format, with status and versions.

File structure, schema, configuration, states, triggers, security and portability. Draft.

The signed message format agents use to talk to each other. Draft.

OpenAPI document of the Daemon API, for generating clients. Reference: Daemon API.