Architecture

AI Agent Framework separates workflow composition from provider integration and agent execution.

Components

Component	Responsibility
Runnables	composable execution units
Prompt templates	prompt formatting
Output parsers	conversion from model text to typed output
Model adapters	provider integration
Agent runtime	iterative execution and tool handling

Chain Execution

input -> runnable stages -> structured output

Chains run a known sequence of stages.

Agent Execution

input -> model -> tool calls or final response -> repeat

Agents can continue across multiple steps until they return a final response or reach maxSteps.

Runtime Behavior

The runtime validates tool arguments, preserves tool-call message ordering, and raises explicit errors for provider failures, invalid tool usage, parser failures, and step-limit exhaustion.

Detailed loop behavior:

1. a system message is prepended with current memory context
2. model output is interpreted into instructions
3. optional instruction passes run
4. tool instructions execute and tool results are appended to:
   • message history (role=tool)
   • memory (type=tool-result)
5. return instruction ends the loop

Tracing:

• each step creates a span with startedAt, endedAt, and step metadata
• spans are stored on agent state for downstream observability/instrumentation

For exact runtime guarantees, see the upstream Core Contract.