Agents

An agent is a definition, not a process: a name, a backend (which CLI or engine runs it), a model, a system prompt, which Tools and Secrets it’s allowed to touch. Launching it starts a run; the definition itself just sits in ~/.adi/mono/agents/ until you do. adi-agent, the environment’s own default agent, is a definition like any other — nothing about it is special-cased except that it starts with every tool enabled.

Why it exists

Point a coding CLI at a terminal and it forgets everything the moment you close the tab. An agent definition survives that: the same prompt, the same tools, the same model choice, run again tomorrow or by a trigger at 3am, with every run’s outcome kept (adi-mono agents runs) so “did it work” is a question you ask the store instead of hunting logs.

Backends

Backend	What it is
`pty:claude` / `pty:codex`	The CLI in a live terminal session, on your own subscription login
`process:claude`	The Claude CLI headless, one `--print` turn
`process:codex`	The Codex CLI headless, with later replies resuming the same thread
`harness:claude-sdk`	`claude --print` under ADI’s own harness — a turn cap, scoped tools
`harness:adi`	ADI’s own agent loop, multi-provider — needs a provider API key

Do it

List: adi-mono agents list. Panel: /agents.

adi-mono agents run my-agent
adi-mono agents runs --since 6h          # every run, newest first, with how it ended
adi-mono agents questions                # what's waiting on you to answer

Nothing is on by defaultEnterprise interest

A new agent starts with an empty grant in both directions. A tool exists on the machine — it isn’t usable by an agent until that agent’s definition ticks it on (bin_tools); an engine’s own built-in surface (file access, web, a second task tracker) exists in the CLI — it isn’t callable until named in that agent’s Allowed tools. An agent that names nothing runs on ADI’s own Bash/Await/Ask alone. Widening either list is a deliberate, per-agent, auditable act — not a machine-wide switch. adi-agent, this environment’s own default, is the one deliberate exception: it ships with every tool ticked on, because its job is operating the whole environment.

When a run can’t decideExpert area

A run can’t sit and wait for a person — a turn ends like any other. When it hits a genuine fork it calls Ask, which writes the question down and ends the turn with nothing held open; the answer arrives as a new message with the whole transcript still there. adi-mono agents answer settles it from anywhere. An agent nobody would see the question from — a trigger’s agent, a 3am sweep — is marked unattended, and Ask refuses so the run decides for itself instead of stopping silently until morning.

Two numbers cap how many runs may be live at once: a global ceiling and, per project, a narrower one — adi-mono agents limit. Automatic launches (a trigger, a queued chat turn) are refused once a cap is full; a human overrides it with --force.