generated: '2026-07-21'
method: searched
source: https://docs.symbolica.ai
authentication:
  style: platform-api-key
  detail: Single Agentica Platform API key via environment variable (see authentication/symbolica-authentication.yml).
integration_model:
  style: sdk-first
  detail: >-
    Direct SDK integration ("Un-MCP") rather than a REST API. Agents and agentic functions
    are defined in Python or TypeScript; the SDK executes them against the Symbolica Platform
    (sandboxed REPL + inference). Python and TypeScript objects/modules are "warped" by
    reference or by value into agent REPLs.
streaming:
  supported: true
  detail: >-
    Responses stream as chunks. Python uses StreamLogger; TypeScript uses a listener
    callback. Chunk.type values include reasoning, output_text, and usage. Token usage
    (including cached and reasoning tokens) is exposed via ResponseUsage.
  docs: https://docs.symbolica.ai/concepts/logging_and_streaming/python.md
mcp_interop:
  style: un-mcp
  detail: >-
    The Agentica Python SDK consumes external MCP servers (local or remote) by turning MCP
    tools back into regular functions for use inside the agent REPL. Symbolica does not
    publish a first-party hosted MCP server; MCP is an inbound compatibility layer.
  docs: https://docs.symbolica.ai/concepts/unmcp.md
error_handling:
  operational_vs_agent: >-
    Operational errors (platform/network/inference) are retried automatically with
    exponential backoff and jitter; only unrecoverable ones reach code. Agent errors are
    intentionally raised by the agent from business logic. All SDK errors inherit from
    AgenticaError.
  docs: https://docs.symbolica.ai/guides/operational-errors
  see: errors/symbolica-error-codes.yml
retries:
  strategy: exponential-backoff-with-jitter
  scope: transient operational failures (inference rate limits, network, service unavailable)
  handled_by: platform
token_controls:
  max_tokens: Configurable per agent/agentic function; MaxTokens supports per_invocation, per_round, and rounds.
  reasoning_effort: "none | minimal | low | medium | high | xhigh (thinking budget for reasoning models)"
  cache_ttl: Controls Anthropic prompt cache duration.
human_in_the_loop:
  supported: true
  docs: https://docs.symbolica.ai/guides/human-in-the-loop
idempotency:
  supported: false
  note: No idempotency-key contract is documented for the SDK/platform surface.
versioning:
  scheme: semver
  see: lifecycle/symbolica-lifecycle.yml
notes: >-
  Cross-links: authentication/, errors/, lifecycle/. No pagination or REST rate-limit-header
  conventions apply because there is no documented REST resource surface.