CaniFriend

When you can't be there, your neighborhood can.

Comment

CaniFriend

CaniFriend is an AI neighborhood care network for pets. When a pet parent cannot get home, CaniFriend turns a natural-language need into a trusted, governed care handoff: understand the pet's context, rank familiar carers, propose a plan, require owner approval, coordinate the care mission, verify completion, and write the outcome back to the pet's longitudinal record.

The problem

Pet care emergencies are often small but urgent: a late meeting, delayed flight, unexpected hospital visit, or traffic jam can leave a pet without dinner, water, medication reminders, or a familiar human check-in. The people who could help may already be nearby, but coordinating the right person with the right instructions and permissions is fragmented across messages and memory.

Our 90-second judge scenario is intentionally simple: “I can't get home tonight. Make sure Pika gets dinner.”

Who it is for

CaniFriend is entered in the Good Neighbor Agents track. It is for pet parents and trusted people around them - neighbors, friends, family, community carers, and local pet-support networks.

What the agent does

The CaniFriend Strands Care Agent handles the coordination work end to end:

  1. Retrieves Pika's bounded pet-care context.
  2. Finds and ranks trusted carers based on authorization, availability, familiarity, and care history.
  3. Creates a proposed care plan.
  4. Stops at an explicit Owner Approval Required gate.
  5. After approval, deterministic policy transitions the plan to an authorized care mission.
  6. The carer completes the physical task and supplies proof.
  7. The verified outcome is recorded as a CAIOS CareEvent for future continuity.

In the demo, Neighbor A is selected as the strongest trusted match, with a 7:00 PM target and CaniBowl Chicken 100g dinner instructions.

Safety by architecture

Our governing principle is: LLM recommends. Policy authorizes. Human approves consequential handoff.

The model cannot authorize home access, approve the owner handoff, accept on behalf of a carer, claim physical care completion, or write an authoritative outcome by itself. Those transitions are controlled by explicit human and deterministic policy boundaries.

This matters because a useful community agent should not maximize autonomy at the expense of trust.

Architecture

Pet Owner -> CaniFriend Web/BFF -> Amazon Bedrock AgentCore Runtime -> Strands Care Agent -> Safe Tool Layer -> Human Owner Approval Gate -> Deterministic Care Mission -> verified outcome -> CAIOS System of Record.

The Safe Tool Layer includes bounded tools such as get_pet_context, find_trusted_carers, create_care_plan, approval/acceptance transitions, task completion, and outcome recording.

AgentCore Memory and agent context are separated from CAIOS, which remains the authoritative pet system of record.

Real AWS execution proof

The hackathon build has a working AWS deployment in us-west-2:

  • ARM64 container built with AWS CodeBuild
  • container stored in Amazon ECR
  • Amazon Bedrock AgentCore Runtime status: READY
  • Strands Agents live invocation returned HTTP 200
  • live result selected Neighbor A for Pika's 19:00 dinner
  • owner approval remained required
  • Amazon CloudWatch captured the successful runtime session

The judge UI includes a Recording Mode and an agent trace so the safety boundary and state transitions can be understood without exposing private chain-of-thought.

Why CaniFriend is different

CaniFriend is not another sitter marketplace. It treats existing human trust as infrastructure and uses an agent to coordinate it safely. The core loop is:

Need -> Reason -> Match -> Human Approval -> Coordinate -> Care -> Verify -> Remember

The broader CAIOS ecosystem gives this a durable role: CAIOS remembers the pet; CARLI can observe and interact with the physical environment; CaniFriend brings a trusted human when physical help is needed.

Impact

A neighborhood-scale care agent can reduce the coordination cost of asking for help while preserving owner control. The same governed handoff pattern can later extend from dinner and water to familiar check-ins, grooming pickups, clinic transport, temporary care, and community pet-support programs.

One pet. One missed dinner. One trusted neighbor. One agent completes the handoff.

Built With

Share this project:

Updates

Submission history