Agents

Agents are the core building block of AgentBackend. Each agent wraps an LLM with tools, memory, guardrails, and an orchestration strategy. Choose from five orchestration types to match your use case.

Agent Types

Single Agent

single

Standalone agent for straightforward tasks like Q&A, content generation, or simple data lookups.

Chain

chain

Sequence of agents where the output of one becomes the input of the next. Ideal for multi-step pipelines like research-then-summarize.

Supervisor

supervisor

Manager agent that delegates sub-tasks to specialized worker agents and synthesizes results.

Society of Mind

society_of_mind

Multiple agents collaborate, debate, and refine outputs through structured dialogue. Best for generating diverse perspectives.

Workflow

workflow

Directed graph with conditional branching, parallel execution, and merge points. For complex business processes.

Configuration Fields

Every agent is defined by a JSON configuration object. Here is the full schema with all available fields:

agent-config.json
{
  "name": "string",
  "type": "single | chain | supervisor | society_of_mind | workflow",
  "model": "openai/gpt-4o-mini",
  "instructions": "System prompt",
  "description": "Agent description",
  "tools": [{"name": "web_search", "type": "catalog"}],
  "deployed": true,
  "memory": {"type": "semantic"},
  "guardrails": {"moderation": true, "pii_filter": true},
  "max_turns": 10,
  "planning": false,
  "reasoning": false
}
  • nameDisplay name for your agent.
  • typeOrchestration pattern — see Agent Types above.
  • modelLLM model ID in provider/model format. Defaults to vertex_ai/gemini-3-flash-preview.
  • instructionsSystem prompt that defines agent behavior, tone, and constraints.
  • toolsArray of tool objects. Each tool has a name and type (catalog or custom).
  • memoryMemory configuration. Semantic memory enables cross-session recall.
  • guardrailsSafety filters — moderation for harmful content, pii_filter for personal data.
  • max_turnsMaximum number of agent turns per run. Prevents runaway loops.
  • planningEnable multi-step planning before execution. Useful for complex tasks.
  • reasoningEnable chain-of-thought reasoning. Improves accuracy at the cost of speed.

Supported Models

Model IDProviderSpeedCost
vertex_ai/gemini-3-flash-previewGoogleFastCheap (default)
openai/gpt-4oOpenAIMediumMedium
openai/gpt-4o-miniOpenAIFastCheap
anthropic/claude-sonnet-4-20250514AnthropicMediumMedium
anthropic/claude-haiku-4-5-20251001AnthropicFastCheap

Powered by LiteLLM — all LiteLLM-supported models are available. Use the provider/model format for the model ID.

Tools Catalog

28 built-in tools organized by category. Add any tool to your agent with {"name": "tool_name", "type": "catalog"}.

Math & Utility

calculatorcurrent_datetimegenerate_uuidrandom_generatorunit_converterencoderjson_tooltext_toolcode_execute

Search & Web

web_searchweb_scrapernews_searchread_documentyoutube_transcript

Communication

send_emailsend_smsslack_send

Data & Storage

google_sheetsdatabase_querytenant_querytenant_schema_infogoogle_calendar

AI & Media

generate_imagetranslate

Developer

githubstripe

Location

google_mapsweather

Presets

Presets are pre-configured agent templates that bundle tools, model settings, and modes for common use cases.

researcher

Web search + scraper, planning mode, reasoning

coder

Code execution, GitHub integration

reviewer

Code review, structured output

assistant

General purpose, web search

Built-in Agents

AgentBackend also provides a built-in Database Agent for designing and managing Data Store schemas. See Data Store docs for details.