Swarms API Documentation - Build AI Agents & Multi-Agent Systems

Overview

This tutorial shows how to run Claude Fable 5 — Anthropic’s most capable model, positioned above the Opus family — on the Swarms API. Fable 5 is built for the hardest reasoning and long-horizon agentic work: multi-step analysis, deep synthesis, and debates where agents must genuinely engage with each other’s arguments. Using it requires exactly two things in your agent config: set model_name to anthropic/claude-fable-5, and leave temperature as None (the model rejects numeric sampling parameters). Everything else about your existing agents and swarms stays the same. The examples below walk through the full range of agent types — a standalone single agent, a sequential pipeline, a concurrent specialist fan-out, a director-led hierarchical swarm, a multi-agent group chat, and a cost-efficient mixed-model pattern that reserves Fable 5 for the synthesis step.

What This Example Shows

The correct model name to use (anthropic/claude-fable-5)
Why temperature must be None for this model — never a number
A single agent on Claude Fable 5
A SequentialWorkflow research pipeline
A ConcurrentWorkflow fan-out across specialist agents
A HierarchicalSwarm with an auto-generated director
A GroupChat debate between opposing analysts
A mixed-model pattern: Fable 5 as the aggregator, cheaper models as workers

Claude Fable 5 is Anthropic’s most capable generally available model, sitting above the Opus family. It is built for the hardest reasoning and long-horizon agentic work, and it is wired into every multi-agent primitive on the Swarms platform. Switching an existing agent to Fable 5 is a one-line change in your agent config.

Claude Fable 5 does not accept the temperature parameter (or top_p). The field must be None — either leave it out of your agent_config entirely or set it explicitly to None, and the Swarms API will omit it from the upstream call. Passing any numeric value is rejected by Anthropic with a 400 error.

Step 1: Get Your API Key

Visit https://swarms.world/platform/api-keys
Sign in or create an account
Generate a new API key
Set it as an environment variable:

export SWARMS_API_KEY="your-api-key-here"

Step 2: Install the Swarms Python Client

pip install swarms-client python-dotenv

Step 3: Create the Client

Every example below reuses this client:

import json
import os

from dotenv import load_dotenv
from swarms_client import SwarmsClient

load_dotenv()

client = SwarmsClient(
    api_key=os.getenv("SWARMS_API_KEY"),
    base_url="https://api.swarms.world",
    timeout=1000,
)

Fable 5 reasons deeply before answering, so individual calls can take noticeably longer than smaller models — especially on hard tasks. Keep the client timeout generous.

Single Agent on Claude Fable 5

The minimal case. Note that temperature is explicitly None — this is the only sampling configuration Fable 5 accepts.

result = client.agent.run(
    agent_config={
        "agent_name": "Quant Research Analyst",
        "description": "A quantitative analyst specializing in derivatives pricing and risk.",
        "system_prompt": (
            "You are a senior quantitative research analyst. You reason carefully "
            "through multi-step financial problems, state your assumptions explicitly, "
            "and show the intermediate calculations that support your conclusion."
        ),
        "model_name": "anthropic/claude-fable-5",
        "temperature": None,  # required: Fable 5 rejects numeric temperature
        "max_loops": 1,
        "max_tokens": 8192,
    },
    task=(
        "A 6-month European call on a non-dividend stock trades at $7.20. "
        "Spot is $100, strike is $105, and the risk-free rate is 4%. "
        "Back out the implied volatility and sanity-check it against a 25% "
        "historical vol. Is the option rich or cheap?"
    ),
)

print(json.dumps(result, indent=4))

What changed from a typical agent call:

Field	Before	With Fable 5
`model_name`	`gpt-4.1`, `anthropic/claude-opus-4-8`, etc.	`anthropic/claude-fable-5`
`temperature`	`0.5` (or any float)	`None` — or omit the field entirely
Everything else	—	unchanged

SequentialWorkflow: Research → Analysis → Report

Agents run one after another; each receives the previous agent’s output. Fable 5’s long-horizon reasoning makes it a strong fit for every stage of a pipeline like this.

result = client.swarms.run(
    name="Fable 5 Research Pipeline",
    description="Researcher gathers facts, analyst interprets them, writer produces the deliverable.",
    swarm_type="SequentialWorkflow",
    task="Produce a briefing on the current state of solid-state battery commercialization.",
    max_loops=1,
    agents=[
        {
            "agent_name": "Researcher",
            "description": "Gathers and organizes the relevant facts.",
            "system_prompt": (
                "You are a research specialist. Collect the key facts, players, and "
                "timelines relevant to the task. Output a structured fact sheet."
            ),
            "model_name": "anthropic/claude-fable-5",
            "temperature": None,
            "max_loops": 1,
            "max_tokens": 4096,
        },
        {
            "agent_name": "Analyst",
            "description": "Interprets the research and extracts implications.",
            "system_prompt": (
                "You are an industry analyst. Take the fact sheet you are given and "
                "identify the three most important implications, with reasoning."
            ),
            "model_name": "anthropic/claude-fable-5",
            "temperature": None,
            "max_loops": 1,
            "max_tokens": 4096,
        },
        {
            "agent_name": "Report Writer",
            "description": "Turns the analysis into a polished briefing.",
            "system_prompt": (
                "You are a professional writer. Turn the analysis you receive into a "
                "concise executive briefing with a clear bottom line up front."
            ),
            "model_name": "anthropic/claude-fable-5",
            "temperature": None,
            "max_loops": 1,
            "max_tokens": 4096,
        },
    ],
)

print(json.dumps(result, indent=4))

ConcurrentWorkflow: Specialist Fan-Out

All agents receive the same task in parallel and answer independently — ideal when you want several expert perspectives at once.

result = client.swarms.run(
    name="Fable 5 Deal Review Panel",
    description="Three specialists review the same acquisition target concurrently.",
    swarm_type="ConcurrentWorkflow",
    task=(
        "Evaluate the acquisition of a 200-person B2B SaaS company with $30M ARR, "
        "110% net revenue retention, and 18 months of runway. Flag the biggest risk "
        "from your specialty's point of view."
    ),
    max_loops=1,
    agents=[
        {
            "agent_name": "Financial Diligence",
            "description": "Reviews the financial profile of the target.",
            "system_prompt": "You are an M&A financial diligence expert. Assess unit economics, burn, and valuation risk.",
            "model_name": "anthropic/claude-fable-5",
            "temperature": None,
            "max_loops": 1,
            "max_tokens": 4096,
        },
        {
            "agent_name": "Technical Diligence",
            "description": "Reviews the technology and engineering organization.",
            "system_prompt": "You are a technical diligence expert. Assess architecture risk, technical debt, and key-person risk.",
            "model_name": "anthropic/claude-fable-5",
            "temperature": None,
            "max_loops": 1,
            "max_tokens": 4096,
        },
        {
            "agent_name": "Legal Diligence",
            "description": "Reviews contracts, IP, and compliance exposure.",
            "system_prompt": "You are a legal diligence expert. Assess contract concentration, IP ownership, and regulatory exposure.",
            "model_name": "anthropic/claude-fable-5",
            "temperature": None,
            "max_loops": 1,
            "max_tokens": 4096,
        },
    ],
)

print(json.dumps(result, indent=4))

HierarchicalSwarm: Director + Workers

You define only the workers — the framework auto-generates a director that decomposes the task, routes subtasks to each worker, and synthesizes the final answer.

result = client.swarms.run(
    name="Fable 5 Markets Swarm",
    description="Director coordinates an ETF analyst and a stocks analyst.",
    swarm_type="HierarchicalSwarm",
    task=(
        "Compare the outlook for the SPY ETF and NVDA stock for the next quarter. "
        "Highlight the strongest signal for each."
    ),
    max_loops=1,
    agents=[
        {
            "agent_name": "ETF Analyst",
            "description": "Analyzes broad-market and sector ETFs.",
            "system_prompt": (
                "You are an ETF analyst. Given a ticker, summarize the fund's "
                "exposure, recent flows, and near-term outlook in under 200 words."
            ),
            "model_name": "anthropic/claude-fable-5",
            "temperature": None,
            "role": "worker",
            "max_loops": 1,
            "max_tokens": 4096,
        },
        {
            "agent_name": "Stocks Analyst",
            "description": "Analyzes individual equities.",
            "system_prompt": (
                "You are an equity analyst. Given a ticker, summarize the company's "
                "fundamentals, catalysts, and near-term outlook in under 200 words."
            ),
            "model_name": "anthropic/claude-fable-5",
            "temperature": None,
            "role": "worker",
            "max_loops": 1,
            "max_tokens": 4096,
        },
    ],
)

print(json.dumps(result, indent=4))

GroupChat: Opposing Analysts Debate

Agents discuss the task together and can respond to each other’s arguments. Fable 5 is notably strong at pushing back on weak reasoning, which makes debates substantive rather than agreeable.

result = client.swarms.run(
    name="Fable 5 Investment Debate",
    description="A bull and a bear debate the thesis; a moderator keeps score.",
    swarm_type="GroupChat",
    task="Debate: should a long-term portfolio overweight AI infrastructure stocks at current valuations?",
    max_loops=2,
    agents=[
        {
            "agent_name": "Bull",
            "description": "Argues the optimistic case.",
            "system_prompt": "You argue the strongest honest bull case. Engage directly with the bear's specific points.",
            "model_name": "anthropic/claude-fable-5",
            "temperature": None,
            "max_loops": 1,
            "max_tokens": 4096,
        },
        {
            "agent_name": "Bear",
            "description": "Argues the pessimistic case.",
            "system_prompt": "You argue the strongest honest bear case. Engage directly with the bull's specific points.",
            "model_name": "anthropic/claude-fable-5",
            "temperature": None,
            "max_loops": 1,
            "max_tokens": 4096,
        },
        {
            "agent_name": "Moderator",
            "description": "Weighs both sides and summarizes.",
            "system_prompt": "You are a neutral moderator. After each round, identify which arguments landed and which were rebutted.",
            "model_name": "anthropic/claude-fable-5",
            "temperature": None,
            "max_loops": 1,
            "max_tokens": 4096,
        },
    ],
)

print(json.dumps(result, indent=4))

Cost Pattern: Fable 5 Aggregator, Cheaper Workers

Fable 5 is priced above the Opus tier, so a common production pattern is MixtureOfAgents with inexpensive workers doing the breadth work and Fable 5 doing the final synthesis, where its reasoning matters most.

result = client.swarms.run(
    name="Mixed-Model Mixture of Agents",
    description="Cheap workers draft perspectives; Fable 5 synthesizes the final answer.",
    swarm_type="MixtureOfAgents",
    task="What are the best practices for securing a multi-tenant Kubernetes cluster?",
    max_loops=1,
    agents=[
        {
            "agent_name": "Network Security Worker",
            "description": "Covers network policy and segmentation.",
            "system_prompt": "You are a network security engineer. Cover network policies, segmentation, and ingress hardening.",
            "model_name": "claude-haiku-4-5-20251001",
            "max_loops": 1,
            "max_tokens": 2048,
        },
        {
            "agent_name": "Identity Worker",
            "description": "Covers RBAC and workload identity.",
            "system_prompt": "You are an identity engineer. Cover RBAC, service accounts, and workload identity.",
            "model_name": "claude-haiku-4-5-20251001",
            "max_loops": 1,
            "max_tokens": 2048,
        },
        {
            "agent_name": "Synthesizer",
            "description": "Merges worker outputs into one authoritative answer.",
            "system_prompt": (
                "You are the lead security architect. Merge the specialist inputs into a "
                "single prioritized checklist, resolving any conflicts between them."
            ),
            "model_name": "anthropic/claude-fable-5",
            "temperature": None,
            "max_loops": 1,
            "max_tokens": 8192,
        },
    ],
)

print(json.dumps(result, indent=4))

Using Fable 5 in Other Swarm Types

Every other swarm architecture takes the same model_name. Drop "model_name": "anthropic/claude-fable-5" (with temperature left as None) into any agent inside any of these swarm configs:

SequentialWorkflow
ConcurrentWorkflow
AgentRearrange
MixtureOfAgents
GroupChat
MajorityVoting
CouncilAsAJudge
MultiAgentRouter
HeavySwarm
LLMCouncil
DebateWithJudge
BatchedGridWorkflow
RoundRobin
PlannerWorkerSwarm
auto

Common Pitfalls

400 error mentioning temperature, top_p, or top_k

You passed a numeric sampling parameter. Fable 5 removed temperature, top_p, and top_k entirely. Set temperature to None (or delete the field) in every agent config that uses anthropic/claude-fable-5 — the Swarms API strips None fields before calling Anthropic.

Requests take longer than with other models

Expected. Fable 5 always reasons before answering and hard tasks can run for several minutes. Raise your client timeout, and prefer swarm architectures that parallelize independent work (ConcurrentWorkflow, MixtureOfAgents) over long sequential chains when latency matters.

A request returns empty content with a refusal stop reason

Fable 5 ships with additional safety classifiers and may decline certain requests (notably deep cybersecurity and research-biology content) that other models answer. If your workload lives near those domains, route it to anthropic/claude-opus-4-8 instead.

Where do I find pricing?

Fable 5 pricing follows Anthropic’s published rates (

10/

50 per million input/output tokens) and is reflected in the Swarms cost-tracking endpoints. See the pricing page for current rates.

Next Steps

Browse the Multi-Agent Architectures catalog for more swarm types
Read the Single Agent Overview for the full agent config surface
See the Claude Opus 4.8 example for the lower-cost Anthropic tier with the same no-temperature rule
See Streaming to stream Fable 5 tokens to your client in real time

Getting Started

Agent Capabilities

Multi-Agent Architectures

Finance & Investing

Healthcare & Life Sciences

Legal, M&A & Diligence

Customer Operations

Defense, Gov & Supply Chain

Marketing, Content & Founder Tools

API Utilities

Migration Guides

Resources

Production

Compatibility & Benchmarks

Tutorials & Integrations

Claude Fable 5

Overview

What This Example Shows

Step 1: Get Your API Key

Step 2: Install the Swarms Python Client

Step 3: Create the Client

Single Agent on Claude Fable 5

SequentialWorkflow: Research → Analysis → Report

ConcurrentWorkflow: Specialist Fan-Out

HierarchicalSwarm: Director + Workers

GroupChat: Opposing Analysts Debate

Cost Pattern: Fable 5 Aggregator, Cheaper Workers

Using Fable 5 in Other Swarm Types

Common Pitfalls

Next Steps

​Overview

​What This Example Shows

​Step 1: Get Your API Key

​Step 2: Install the Swarms Python Client

​Step 3: Create the Client

​Single Agent on Claude Fable 5

​SequentialWorkflow: Research → Analysis → Report

​ConcurrentWorkflow: Specialist Fan-Out

​HierarchicalSwarm: Director + Workers

​GroupChat: Opposing Analysts Debate

​Cost Pattern: Fable 5 Aggregator, Cheaper Workers

​Using Fable 5 in Other Swarm Types

​Common Pitfalls

​Next Steps

Overview

What This Example Shows

Step 1: Get Your API Key

Step 2: Install the Swarms Python Client

Step 3: Create the Client

Single Agent on Claude Fable 5

SequentialWorkflow: Research → Analysis → Report

ConcurrentWorkflow: Specialist Fan-Out

HierarchicalSwarm: Director + Workers

GroupChat: Opposing Analysts Debate

Cost Pattern: Fable 5 Aggregator, Cheaper Workers

Using Fable 5 in Other Swarm Types

Common Pitfalls

Next Steps