Skip to content
Open
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
53 changes: 18 additions & 35 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -5,41 +5,24 @@ A minimalist framework for ambient, always-on agents.

## The idea

**Agents should wait for events.** An agent built on a request loop knows
nothing between calls: when a deadline passes or a task completes,
something else has to notice and invoke it. Ambion inverts the loop:

- Agents wait in a session, and every event enters it as a message — a
person speaking, a person walking in, a timer firing, a task completing,
a system reporting.
- A message activates exactly the agents it concerns. The rest stay at
rest.
- Cost follows events. An idle session costs nothing.

**The agent is the unit of context engineering, and the ownership
boundary.** Building one agent is already this work: choosing its tools
and shaping their responses, disclosing context progressively, adding
guardrails and completion checks. The techniques hold because they serve
one domain. In a shared context window they collide — every change lands
in every domain's context — and the shared agent settles at a local
maximum, where no team can improve its domain without degrading another's.
Ambion draws the boundary at the agent:

- One team owns one agent whole: domain, tools, instructions, model,
evals.
- Inside one agent every technique serves the same domain, so the
techniques compose.
- A team improves its agent on its own schedule.

**Every agent that keeps growing arrives at multi-agent collaboration.
Ambion starts there.** When one context window stops holding the work, the
fix is subagents — a team, inside one engine, with no owners and no record
of how it works together. Ambion makes the team first-class:

- Agents with owners, working in a workspace built for collaboration.
- An agent reads what a colleague did and calls that colleague in by name.
- The platform provides what they share: the session and the record. The
complexity lives in the composition.
**Agents should wait for events.** An agent on a request loop knows nothing
between calls. When a deadline passes or a task completes, something else
has to notice and invoke it. Ambion inverts the loop. Agents wait in a
session, every event enters it as a message, and a message activates the
agents it concerns. An idle session costs nothing.

**The agent is the ownership boundary.** Choosing tools, shaping their
responses, disclosing context progressively, adding guardrails: these
techniques hold because they serve one domain. In a shared context window
every change lands in every domain's context. The agent settles where no
team can improve its domain without degrading another's. One team owns one
agent whole — domain, tools, instructions, model, evals.

**Every agent that keeps growing arrives at multi-agent collaboration.**
When one context window stops holding the work, the fix is subagents: a
team inside one engine, with no owners. Ambion makes the team first-class.
An agent reads what a colleague did and calls that colleague in by name.
The platform provides what they share: the session and the record.

## The activation rule

Expand Down
Loading