LLM-based Agents

Agents in Pika run on Amazon Bedrock. They stream answers, call tools with typed inputs, and operate under clear access rules.

How agents work

• Base prompt defines role and guardrails.
• Tool references give the agent capabilities (each tool has an input/output schema).
• Access rules and rollout control who can use what, and where.
• Caching can be enabled per agent for performance and testing.

Chat app, agent and tool definitions live in a centralized registry (DynamoDB), so you can evolve capabilities without redeploying the Pika infrastructur and future-proof your agents as new technologies come online (e.g. MCP).

Tool calling with schemas

Tools run in Lambda and declare JSON function schemas. The agent validates inputs and returns typed results, improving reliability and transparency.

Streaming responses

Users see answers as they’re generated. The platform updates history and usage metrics when the turn completes.

Bring your knowledge

Augment agents with your data sources and domain tools. Use the registry to experiment safely while keeping a clear audit trail of changes.