CAIOS Care Agent

CAIOS Care Agent

CAIOS Care Agent is a governed AI follow-up copilot for veterinary and companion-animal care workflows. It helps professionals and pet parents turn scattered longitudinal context into a safe, auditable next action without giving an autonomous diagnosis or silently changing treatment.

The problem

Follow-up care often depends on information that lives across owner messages, home observations, prior recommendations, clinic notes, reminders, and outcomes. The repetitive work is not only collecting that context — it is deciding what matters now, whether an action is low-risk, whether owner approval is required, and when the correct action is to stop automation and escalate to a veterinary professional.

Who it is for

We are entering the Professional Agents track. The primary user is a veterinary professional or care team handling follow-up, continuity, and owner communication. A companion-facing surface can also present approved low-risk actions to the pet parent.

What the agent does

The agent follows an explicit governed state machine:

observe -> classify -> propose -> authorize -> act -> record -> evaluate

Each proposed action is classified into one of three authority classes:

1. SAFE_AUTONOMOUS — low-risk reminders or context gathering only.
2. OWNER_APPROVAL — explicit owner confirmation is required before execution.
3. PROFESSIONAL_ESCALATION — automation stops and surfaces veterinary or emergency guidance.

The system is designed so that an LLM does not own medical authority. Deterministic red-flag checks happen before agent action, professional escalation cannot be overridden by the model, and every action records its context, reason, authorization, execution state, and outcome.

Why Strands Agents

For this hackathon, Strands Agents is used as the agent orchestration adapter over CAIOS's provider-neutral Care Agent Runtime. The domain policy remains portable, while Strands handles the agent/tool interaction layer. This lets us demonstrate genuine agent execution without coupling pet-safety policy to one model or runtime provider.

End-to-end demo flow

The demo uses synthetic pet data and shows a veterinary follow-up scenario:

1. A new home observation arrives after a prior consultation.
2. The agent retrieves longitudinal context and classifies risk.
3. It proposes the next action with provenance.
4. The runtime decides whether the action is safe-autonomous, requires owner approval, or must escalate to a professional.
5. If authorized, a bounded tool action is executed.
6. The result is recorded as an auditable outcome for the next follow-up cycle.

A second scenario demonstrates a red-flag condition where the agent is not allowed to continue automation and instead escalates.

Safety and governance

CAIOS Care Agent intentionally does not autonomously diagnose, prescribe, change medication, or override veterinary judgment. The architecture uses:

• deterministic safety checks before model execution;
• explicit action authority classes;
• human-in-the-loop approval for consequential actions;
• durable idempotency and replay protection;
• bounded error messages that do not leak secrets;
• immutable provenance linking context, recommendation, decision, action, and outcome;
• recovery semantics for ambiguous external side effects.

Architecture

The hackathon build is an isolated public demonstration layer. It does not expose CAIOS private production code or proprietary longitudinal data.

Synthetic pet context -> deterministic safety -> Strands Agent adapter -> CAIOS Care Agent policy -> approval/escalation -> tool execution -> audit/outcome

The architecture diagram included with the submission shows this authority boundary explicitly.

What makes it different

Most agent demos optimize for autonomy. CAIOS Care Agent optimizes for governed autonomy: the agent can do useful work end to end, but only inside explicit safety and authority boundaries. That distinction matters in veterinary follow-up because the best next action is sometimes a reminder, sometimes an owner decision, and sometimes a deliberate refusal to automate.

Impact

The product is designed around three practical outcomes:

• reduce time spent rebuilding longitudinal context during follow-up;
• improve completeness of home-to-clinic information;
• improve owner adherence and follow-up completion.

The hackathon prototype demonstrates the technical execution layer behind that workflow with synthetic data and auditable actions.

Built With

• amazon-bedrock
• amazon-web-services
• fastapi
• pydantic
• pytest
• python
• strands-agents-sdk