Agent Runtime

The engine that runs your AI. For every message it plans a response, stays inside a strict budget and autonomy limit, and routes any real-world action to the Tool Broker. It holds no credentials and no standing power of its own — it only thinks and routes.

In plain terms

The Agent Runtime is the engine that runs your AI. For every message it plans a response, stays inside a strict budget and autonomy limit, and sends any real-world action to the Tool Broker to be authorized. It has no power of its own — it only thinks and routes.

How this differs

Capabilities— what's turned onTool Broker — what actions are allowedAgent Runtime — how each turn actually runs

How every turn runs

One loop per message · the runtime never skips Verify

Each message runs through the same four stages. The engine plans a response and routes any real action out to be authorized — it has no power to act on its own.

Perceive
Reads the incoming message and the context it's allowed to see — nothing more.
Plan
Decides the smallest set of steps that answers the message, inside the budget and ceiling.
Act
→ Tool Broker authorizes
Any real-world action is handed to the Tool Broker to authorize. The runtime itself touches nothing.
Verify
Checks the result is grounded and in-scope before it returns. If not, it degrades or hands off.

Runtime health right now

Turns handled today

1,420+86 last hour

One loop per message

Median turn latency

3.4swithin budget

Plan → verify, end to end

Active turns now

3live

In flight · violet = AI is acting

Budget consumed

41%of per-turn cap

Median across active turns

Autonomy ceilings

The most a turn can do on its own, by who's asking. A ceiling is a cap — never a blank cheque.

GuestA2 · Draft-review

Anyone not signed in — a patient on the website or phone.

Unattended: Answer questions, look things up, draft a reply for review.

Stops for a human: Cannot change records or send anything on its own — a person confirms first.

OperatorA4 · Auto-logged

Signed-in staff — front desk, clinical, billing.

Unattended: Book within policy, send routine confirmations, update non-sensitive fields — all logged.

Stops for a human: Bulk messages, exports and record changes still stop for approval.

OwnerA5 · Fully autonomous

The practice owner or admin — the highest trust level.

Unattended: The widest set of routine actions can run unattended, fully audited.

Stops for a human: Critical and destructive actions are still gated — autonomy is a cap, not a blank cheque.

Per-turn budget envelope

Every turn runs inside a hard cap. Hit the cap and the turn yields — it can't spin forever.

Model calls3 / 8 calls

Reasoning steps via the AI Gateway

Tool calls2 / 6 calls

Real-world actions via the Tool Broker

Wall-clock5 / 20 s

Time budget before the turn yields

Example envelope for one active turn · model calls leave via the AI Gateway · tool calls via the Tool Broker

Degradation ladder

Under load or trouble, the runtime steps down gracefully instead of failing — it never quietly does more.

Current levelD0

D0Full service
Everything healthy — full planning, routing and tools.
D1Trim non-essential
Drops optional enrichment to stay fast under load.
D2Core only
Answers and safe reads only; defers heavier work.
D3Safe replies
Falls back to approved templates; new actions paused.
D4Human handoff
Hands the conversation to a person; no autonomous action.

Recent turns

Read-only · who asked, on what channel, how many steps, and how it ended. Decisions happen in Approvals.

Turn	Who	Channel	Ceiling	Budget used	Outcome
Answered an insurance-coverage question from approved knowledge. turn_4471 · 3 steps	Guest · web-chat	Web chat	A2 · Draft-review	2 model · 1 tool · 3.1s	executed
Booked a cleaning slot — verified availability, then acted within ceiling. turn_4470 · 4 steps	Operator · Maya R.	Voice	A4 · Auto-logged	3 model · 2 tool · 5.4s	executed
Bulk reminder to 142 patients — routed to Tool Broker, held for approval. turn_4469 · 4 steps	Operator · Front desk	Web chat	A4 · Auto-logged	2 model · 1 tool · 2.0s	awaiting_approval
Drafted and published a new-patients page — change staged for review. turn_4468 · 5 steps	Owner · Dr. Chen	Portal	A5 · Fully autonomous	4 model · 3 tool · 9.2s	executed
Requested a refund the AI cannot perform — routed action was refused. turn_4467 · 4 steps	Guest · SMS	SMS	A2 · Draft-review	2 model · 1 tool · 1.4s	failed

No action is decided here · held turns are authorized in the Tool Broker / Approvals