The production-grade agentic metaframework, built on Pydantic AI.

Keep Pydantic AI's Agent, Tool and RunContext — gain lifecycle hooks, delegation, memory, reasoning patterns, validation loops, RAG, and DAG pipelines, all protocol-driven and swappable.

📘 The Tutorial  ·  Quick Start  ·  Why  ·  Architecture  ·  Features  ·  Docs  ·  Ecosystem  ·  Changelog

Copyright 2026 Firefly Software Foundation · Licensed under the Apache License 2.0

• Why Firefly Agentic?
• Key Principles
• Architecture at a Glance
• Feature Highlights
• The Firefly Ecosystem
• Requirements
• Installation
• 5-Minute Quick Start
• Using in Jupyter Notebooks
• Learn the Framework
• Development
• Contributing
• License

Pydantic AI provides an excellent foundation: type-safe, model-agnostic agents with structured output. But a production agentic system demands far more than a single agent call. You need to orchestrate multi-step reasoning, validate and retry LLM outputs against schemas, manage conversation memory across turns, observe every call with traces and metrics, and run A/B experiments to compare models — all without coupling your domain logic to infrastructure concerns.

fireflyframework-agentic is the production framework built on top of Pydantic AI. It extends the engine with composable layers — from core configuration through agent management, intelligent reasoning, experimentation, and pipeline orchestration — so that every concern has a dedicated, protocol-driven module. You write your business logic; the framework provides the architecture.

What "metaframework" means in practice:

• You keep Pydantic AI's familiar Agent, Tool, and RunContext APIs unchanged.
• The framework wraps them with lifecycle hooks, registries, delegation routers, memory managers, reasoning patterns, validation loops, and DAG pipelines — all optional, all composable, all swappable through Python protocols.
• No vendor lock-in: switch models, swap memory backends, or replace components without touching your agent code.

1. Protocol-driven contracts — Every extension point is defined as a @runtime_checkable Protocol or abstract base class. The framework ships 28 such protocols across every layer — AgentLike, ToolProtocol, GuardProtocol, AgentMiddleware, DelegationStrategy, ReasoningPattern, ValidationRule, StepExecutor, Checkpointer, Chunker, MemoryStore, EmbeddingProtocol, VectorStoreProtocol, plus the new workflow ports (AgentRunner, JournalBackend, ModelSelectionStrategy) and more — so you can swap or extend any component without modifying framework internals.

2. Convention over configuration — Sensible defaults everywhere. FireflyAgenticConfig is a Pydantic Settings singleton that reads from environment variables prefixed with FIREFLY_AGENTIC_ and .env files. One config object governs model defaults, retry counts, token limits, telemetry emission (observability_enabled), strict-cost mode (cost_strict), memory backends, and validation thresholds — override only what you need.

3. Layered composition — Layers with strict top-down dependency flow: Core → Agent → Intelligence → Experimentation → Orchestration. Higher layers depend on lower layers but never the reverse, keeping the dependency graph acyclic and each module independently testable.

4. Optional dependencies — Heavy libraries (chromadb, pinecone, openai, asyncpg) are declared as pip extras ([openai-embeddings], [vectorstores-chroma], [postgres], [all]). The core framework imports them lazily inside factory functions so that you install only what your deployment requires.

Every extension point is a @runtime_checkable protocol. Implement the protocol to create your own components; the framework discovers them via duck typing.

• Agents — FireflyAgent wraps pydantic_ai.Agent with metadata, lifecycle hooks, and automatic registration. AgentRegistry provides singleton name-based discovery. DelegationRouter routes prompts across agent pools via seven strategies (RoundRobinStrategy, CapabilityStrategy, ContentBasedStrategy, CostAwareStrategy, ChainStrategy, FallbackStrategy, WeightedStrategy). A composable middleware stack (MiddlewareChain over AgentMiddleware) wraps every run — LoggingMiddleware is always wired and ObservabilityMiddleware is added when observability_enabled, with PromptGuardMiddleware, OutputGuardMiddleware, CostGuardMiddleware, CacheMiddleware, PromptCacheMiddleware, ExplainabilityMiddleware, ValidationMiddleware, RetryMiddleware, and CircuitBreakerMiddleware available to add. FallbackModelWrapper / run_with_fallback provide automatic model failover, and ResultCache / CacheStatistics back response caching. The @firefly_agent decorator defines an agent in one statement. Five template factories (create_summarizer_agent, create_classifier_agent, create_extractor_agent, create_conversational_agent, create_router_agent) cover common use cases out of the box.

• Tools — ToolProtocol (duck-typed) and BaseTool (inheritance) let you choose your extensibility style. ToolBuilder provides a fluent API for building tools without subclassing. Four guard types (ValidationGuard, RateLimitGuard, SandboxGuard, CompositeGuard) intercept calls before execution (a rejected guard raises ToolGuardError). For human-in-the-loop, mark a tool requires_approval=True: the agent run pauses before executing it and returns a DeferredToolRequests (detected via is_deferred(result)), which you resume with deferred_tool_results= — approving (ToolApproved), denying (ToolDenied), or auto-deciding inline via an approval_handler=. The native deferred-tools types are re-exported from fireflyframework_agentic.tools. Three composition patterns (SequentialComposer, FallbackComposer, ConditionalComposer) build higher-order tools. ToolKit groups tools for bulk registration. Nine built-in tools (calculator, datetime, filesystem, HTTP, JSON, search, shell, text, database) are ready to attach to any agent.

• Prompts — PromptTemplate renders Jinja2 templates with variable validation and token estimation. PromptRegistry maps names to versioned templates. Three composers (SequentialComposer, ConditionalComposer, MergeComposer) combine templates at render time. PromptValidator enforces token limits and required sections. PromptLoader loads templates from strings, files, or entire directories.

• Reasoning — Six pluggable patterns implement AbstractReasoningPattern's template-method loop (_reason → _act → _observe → _should_continue): ReAct (observe-think-act), Chain of Thought (step-by-step), Plan-and-Execute (goal → plan → steps with optional replanning), Reflexion (execute → critique → retry), Tree of Thoughts (branch → evaluate → select), and Goal Decomposition (goal → phases → tasks). All produce structured ReasoningResult with ReasoningTrace. Prompts are slot-overridable. Each pattern's structured output is wrapped in a pydantic-ai output mode — selected per-pattern via output_mode= or framework-wide via the reasoning_output_mode config — "tool" (ToolOutput), "native" (provider structured output), or "prompted" (PromptedOutput, portable to any model). OutputReviewer can validate final outputs. ReasoningPipeline chains patterns sequentially.

• Content — TextChunker splits by tokens, sentences, or paragraphs with configurable overlap; MarkdownChunker chunks structure-aware on Markdown headings. DocumentSplitter detects page breaks and section separators. ImageTiler computes tile coordinates for VLM processing. BatchProcessor runs chunks through an agent concurrently with a semaphore. ContextCompressor delegates to pluggable strategies (TruncationStrategy, SummarizationStrategy, MapReduceStrategy) — ContextCompressor.compress is async. SlidingWindowManager maintains a rolling token-budgeted context window. The [binary]-gated content.binary submodule normalises uploaded files into consumer-ready artifacts: BinaryNormalizer (with BinaryConfig) produces BinaryArtifacts, sniff_media_type detects formats, build_office_converter selects an OfficeConverter (GotenbergConverter, LibreOfficeConverter, NoOpOfficeConverter), and PdfGuard, ImageNormalizer, ArchiveUnpacker, and EmailUnpacker handle PDFs, images, archives, and emails.

• Memory — ConversationMemory stores per-conversation turn history with token-budget enforcement (newest-first FIFO eviction). WorkingMemory provides a scoped key-value scratchpad backed by MemoryStore (InMemoryStore, FileStore, or SQLiteStore). MemoryManager composes both behind a unified API and supports fork() for isolating working memory in delegated agents or pipeline branches while sharing conversation context. create_llm_summarizer builds an LLM-backed history summarizer for long conversations.

• Validation — Five composable rules (RegexRule, FormatRule, RangeRule, EnumRule, CustomRule) feed into FieldValidator and OutputValidator. OutputReviewer wraps agent calls with parse-then-validate retry logic: on failure it builds a feedback prompt and retries up to N times. RubricReviewer adds LLM-as-judge grading against a rubric (RubricReviewer.from_rubric_file). QoS guards (ConfidenceScorer, ConsistencyChecker, GroundingChecker, plus the QoSGuard aggregator returning a QoSResult) detect hallucinations and low-quality extractions before they propagate downstream.

• Pipeline — DAG holds DAGNode and DAGEdge objects with cycle detection and topological sort. PipelineEngine executes nodes level-by-level via asyncio.gather for maximum concurrency, with per-node condition gates, retries, and timeouts. PipelineBuilder offers a fluent API (add_node / add_edge / chain). Step types adapt agents, patterns, and functions to DAG nodes: AgentStep, ReasoningStep, CallableStep, FanOutStep, FanInStep, BranchStep, BatchLLMStep, EmbeddingStep, and RetrievalStep. State reducers (append, extend, merge_dict, replace) merge fan-out results, and control signals (Pause, Send) drive branching and human-in-the-loop pauses. Checkpointer / FileCheckpointer (with CheckpointRecord) persist and resume long runs, and a pluggable audit-log family (AuditLog, FileAuditLog, LoggingAuditLog, OtelAuditLog, QueryableAuditLog over AuditEntry) records execution traces.

• Workflows — @workflow / @subworkflow define a code-defined, deterministic orchestration DSL over your agents — a complement to the declarative pipeline DAG, both living in the Orchestration layer. Compose async primitives — agent(), parallel(), pipeline(), stream(), phase(), human() (human-in-the-loop), map_agents() and log() — inside a WorkflowContext that carries a WorkflowBudget (concurrency, agent-count and token/cost ceilings), a Journal (JournalBackend / FileJournalBackend) for deterministic resume, and a pluggable AgentRunner. FireflyAgentRunner (the default) runs every sub-agent call through a full FireflyAgent (middleware, guards, budget, model fallback); DefaultAgentRunner is the lightweight path. SmartRoutingRunner selects the cheapest capable model via a ModelSelectionStrategy (ComplexityHeuristicStrategy, CostFloorStrategy), and verification helpers (cascade, adversarial_verify, judge_panel, loop_until_dry) add refute-by-default quality gates. See docs/workflows.md.

• Observability — FireflyTracer creates OpenTelemetry spans scoped to agents, tools, and reasoning steps. FireflyMetrics records tokens (total, prompt, completion), latency, cost, errors, and reasoning depth via the OTel metrics API. FireflyEvents emits structured log records. @traced and @metered decorators instrument any function with one line. Native pydantic-ai instrumentation is on by default — rich GenAI-convention spans (and metrics) per model request and tool call, nested under the framework's agent span, with prompt/response content stripped by default for privacy (toggle via native_instrumentation_enabled). The framework emits model and agent telemetry purely through the OpenTelemetry API; the host application owns OTel SDK and exporter configuration. UsageTracker automatically records token usage, cost estimates, and latency for every agent run, reasoning step, and pipeline execution. Cost is computed through a resolver chain (resolve_cost, genai_prices_cost, provider_reported_cost, DEFAULT_RESOLVERS); set cost_strict to raise UnknownModelCostError when no price is found. BudgetGate enforces token/cost budgets per scope (BudgetRule, BudgetMode, BudgetWindow), and a pluggable sink family (LoggingSink, JSONLFileSink, OTelMetricsSink, EventBusSink, CostSink) routes usage records wherever you need them.

• Explainability — TraceRecorder captures every LLM call, tool invocation, and reasoning step as DecisionRecord objects. ExplanationGenerator turns records into human-readable narratives. AuditTrail provides an append-only, immutable log with JSON export for compliance. ReportBuilder produces Markdown and JSON reports with statistics.

• Security — PromptGuard scans inbound prompts for injection and jailbreak patterns; OutputGuard redacts secrets and PII from model output (default_prompt_guard / default_output_guard provide ready-to-use instances). At-rest protection comes from AESEncryptionProvider (behind the EncryptionProvider protocol) and EncryptedMemoryStore, which encrypts MemoryEntry.content while leaving keys, metadata, and timestamps in plaintext. Inbound request authentication and authorization are a hosting concern, not the framework's.

• Resilience — CircuitBreaker trips after a configurable failure threshold and rejects calls with CircuitBreakerOpenError while open, transitioning through CircuitState (closed → open → half-open). CircuitBreakerMiddleware plugs it into the agent middleware chain so a failing model is short-circuited before it drains your budget.

• Storage — StorageBackend abstracts blob/object storage with LocalBackend out of the box; DatabaseStore persists artifacts with leasing (WriteSession, LockToken), a configurable RetryPolicy, and StorageMetadata. Typed errors (StorageUploadError, StorageDownloadError, StorageLeaseError, StorageTransientError, StoreUnavailableError) make failure handling explicit.

• Experiments — Experiment defines variants with model, temperature, and prompt overrides. ExperimentRunner executes all variants against a dataset via an agent_factory callable. ExperimentTracker persists results with optional JSON export. VariantComparator computes latency, output length, and comparison summaries.

• Lab — LabSession manages interactive agent sessions with history. Benchmark runs agents against standardised inputs and reports p95 latency. EvalOrchestrator scores agent outputs with pluggable Scorer functions. EvalDataset loads/saves test cases from JSON. ModelComparison runs the same prompts across multiple agents for side-by-side analysis.

• Evaluation — LLM-as-judge metrics (faithfulness, relevancy, answer correctness, RAGAS, …) and deterministic retrieval metrics (recall@k, MRR, MAP, nDCG, …) for assessing LLM and pipeline outputs. Each metric is a plain function you call directly. Install with pip install "fireflyframework-agentic[evaluation]". See docs/evaluation.md for the full guide.

  Optional developer tooling. fireflyframework_agentic.experiments (A/B experiments) and fireflyframework_agentic.lab (offline evaluation / benchmarking) are leaf modules — nothing in the core imports them and they add no third-party dependencies. Import them only if you run experiments or evaluations; agent-building consumers can ignore them.

• Embeddings — EmbeddingProtocol (duck-typed) and BaseEmbedder (inheritance with auto-batching) provide provider-agnostic text embedding. Eight providers ship out of the box: OpenAI, Azure OpenAI, Cohere, Google, Mistral, Voyage AI, AWS Bedrock, and Ollama (local). EmbedderRegistry manages named instances. Built-in similarity utilities (cosine_similarity, euclidean_distance, dot_product) compare vectors without external dependencies. Configuration via embedding_batch_size, embedding_max_retries, and default_embedding_model.

• Vector Stores — VectorStoreProtocol and BaseVectorStore provide pluggable storage and retrieval with six backends: InMemoryVectorStore (zero-dependency, brute-force cosine), ChromaVectorStore, PineconeVectorStore, QdrantVectorStore, PgVectorVectorStore (Postgres + pgvector), and SqliteVecVectorStore (embedded sqlite-vec). A multi-tenant isolation layer (ScopedVectorStore, TenantScopedVectorStore, plus scope_namespace / parse_scope_namespace helpers) namespaces documents per tenant or scope. Auto-embedding upserts documents without pre-computed vectors. search_text embeds a query string and searches in one call, and SearchFilter narrows results by metadata. Namespace scoping isolates document collections. VectorStoreRegistry manages named instances. EmbeddingStep and RetrievalStep integrate directly into DAG pipelines for retrieval-augmented workflows.

• Studio — moved to its own repository: fireflyframework-agentic-studio. A browser-based visual IDE for building agent pipelines (drag-and-drop canvas, code generation, AI assistant, time-travel debugging). Install with pip install fireflyframework-agentic-studio and launch with firefly studio.

Firefly Agentic is the agentic member of the Firefly Framework — a polyglot platform that brings one cohesive programming model to many runtimes. Each member shares the same firefly-in-the-dark identity, recolored per language.

• PyFly — the Python implementation (Spring-Boot DX, async-native).
• Firefly for Rust — reactive, tokio + axum microservices.
• Firefly Studio — a browser-based visual IDE for building agent pipelines.

Runtime:

• Python 3.13 or later
• Git for cloning the repository
• UV package manager (recommended) or pip

Core dependencies (installed automatically):

• pydantic-ai >=1.99.0 — Agent engine (model calls, tool dispatch, streaming)
• pydantic >=2.10.0 — Data validation and settings
• pydantic-settings >=2.7.0 — Environment-based configuration
• Jinja2 >=3.1.0 — Prompt template engine
• httpx >=0.28.0 — Async HTTP client (built-in HTTP tool, Gotenberg converter)
• OpenTelemetry API >=1.29.0 — Tracing and metrics
• OpenTelemetry SDK >=1.29.0 — Telemetry primitives
• genai-prices >=0.0.1 — LLM pricing data for cost resolution
• markdown-it-py >=3.0 — Structure-aware Markdown chunking
• python-dotenv >=1.0.0 — .env loading for example scripts

Optional dependencies (installed via extras):

• [embeddings] — numpy for fast in-memory vector math
• [openai-embeddings] — openai >=1.0.0 for OpenAI/Azure embeddings
• [vectorstores-chroma] — chromadb >=0.5.0
• [vectorstores-pinecone] — pinecone >=5.0.0
• [vectorstores-qdrant] — qdrant-client >=1.12.0
• [vectorstores-pgvector] — asyncpg >=0.30.0 for Postgres + pgvector
• [vectorstores-sqlite-vec] — sqlite-vec >=0.1.6 for embedded vector search
• [binary] — pypdf, Pillow, pillow-heif, cairosvg, py7zr, extract-msg for content.binary
• [all] — Everything (memory backends, security, all embedding providers, all vector stores, watch, binary)

LLM provider keys (at least one):

• OPENAI_API_KEY for OpenAI models
• ANTHROPIC_API_KEY for Anthropic models
• GEMINI_API_KEY for Google Gemini models
• GROQ_API_KEY for Groq models
• Or any Pydantic AI-supported provider

The interactive installer detects your platform, checks Python and UV, lets you choose extras, and installs everything with progress indicators and verification.

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/fireflyframework/fireflyframework-agentic/main/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/fireflyframework/fireflyframework-agentic/main/install.ps1 | iex

Both installers support non-interactive mode for CI/CD:

# macOS / Linux — install with all extras, no prompts
curl -fsSL https://raw.githubusercontent.com/fireflyframework/fireflyframework-agentic/main/install.sh | bash

# Windows — install with all extras, no prompts
.\install.ps1 -NonInteractive -Extras all

git clone https://github.com/fireflyframework/fireflyframework-agentic.git
cd fireflyframework-agentic
uv sync --all-extras # or: pip install -e ".[all]"

python -c "import fireflyframework_agentic; print(fireflyframework_agentic.__version__)"

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/fireflyframework/fireflyframework-agentic/main/uninstall.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/fireflyframework/fireflyframework-agentic/main/uninstall.ps1 | iex

Or manually remove the cloned directory and its virtual environment.

Create a .env file (or set environment variables):

# Provider API key (Pydantic AI reads these automatically)
OPENAI_API_KEY=sk-...
# ANTHROPIC_API_KEY=sk-ant-...
# GEMINI_API_KEY=...
# GROQ_API_KEY=gsk_...

# Framework settings
FIREFLY_AGENTIC_DEFAULT_MODEL=openai:gpt-4o
FIREFLY_AGENTIC_DEFAULT_TEMPERATURE=0.3

The model string format is "provider:model_name" — e.g. "openai:gpt-4o", "anthropic:claude-sonnet-4-20250514", "google:gemini-2.0-flash". Pydantic AI resolves the matching API key from environment variables automatically. For programmatic credential management (Azure, Bedrock, custom endpoints), pass a Pydantic AI Model object directly to FireflyAgent(model=...) — see the tutorial.

from fireflyframework_agentic.agents import firefly_agent

@firefly_agent(name="assistant", model="openai:gpt-4o")
def assistant_instructions(ctx):
    return "You are a helpful conversational assistant."

from fireflyframework_agentic.tools import firefly_tool

@firefly_tool(name="lookup", description="Look up a term")
async def lookup(query: str) -> str:
    return f"Result for {query}"

Human-in-the-loop: mark a tool @firefly_tool(name=..., requires_approval=True) and the agent run pauses before executing it — run() returns a DeferredToolRequests (detect with is_deferred(result)). Resume with agent.run(message_history=paused.all_messages(), deferred_tool_results=DeferredToolResults(approvals={call_id: True})). Full detail in docs/tools.md.

from fireflyframework_agentic.agents import FireflyAgent
from fireflyframework_agentic.memory import MemoryManager

memory = MemoryManager(max_conversation_tokens=32_000)
agent = FireflyAgent(name="bot", model="openai:gpt-4o", memory=memory)

cid = memory.new_conversation()
result = await agent.run("Hello!", conversation_id=cid)
result = await agent.run("What did I just say?", conversation_id=cid)

from fireflyframework_agentic.reasoning import ReActPattern

react = ReActPattern(max_steps=5)
result = await react.execute(agent, "What is the weather in London?")
print(result.output)

from pydantic import BaseModel
from fireflyframework_agentic.validation import OutputReviewer

class Answer(BaseModel):
    answer: str
    confidence: float

reviewer = OutputReviewer(output_type=Answer, max_retries=2)
result = await reviewer.review(agent, "What is 2+2?")
print(result.output) # Answer(answer="4", confidence=0.99)

from fireflyframework_agentic.pipeline.builder import PipelineBuilder
from fireflyframework_agentic.pipeline.steps import AgentStep, CallableStep

pipeline = (
    PipelineBuilder("my-pipeline")
    .add_node("classify", AgentStep(classifier_agent))
    .add_node("extract", AgentStep(extractor_agent))
    .add_node("validate", CallableStep(validate_fn))
    .chain("classify", "extract", "validate")
    .build()
)
result = await pipeline.run(inputs="Process this document")

from fireflyframework_agentic.embeddings.providers import OpenAIEmbedder
from fireflyframework_agentic.vectorstores import InMemoryVectorStore, VectorDocument

embedder = OpenAIEmbedder(model="text-embedding-3-small")
store = InMemoryVectorStore(embedder=embedder)

# Upsert documents (auto-embedded)
await store.upsert([
    VectorDocument(id="1", text="Python is great for AI"),
    VectorDocument(id="2", text="Rust is fast and safe"),
])

# Search by text
results = await store.search_text("machine learning languages", top_k=1)
print(results[0].document.text)  # Python is great for AI

firefly-agentic works seamlessly in Jupyter notebooks and JupyterLab. Since the framework is async-first, use await directly in notebook cells (Jupyter provides a running event loop automatically).

# From your clone directory
cd fireflyframework-agentic
source .venv/bin/activate # activate the venv created by the installer
pip install ipykernel # install Jupyter kernel support
python -m ipykernel install --user --name fireflyagentic --display-name "Firefly Agentic"
jupyter lab # or: jupyter notebook

Then select the Firefly Agentic kernel when creating a new notebook.

# Cell 1 — configure
import os
os.environ["OPENAI_API_KEY"] = "sk-..." # or set in .env
os.environ["FIREFLY_AGENTIC_DEFAULT_MODEL"] = "openai:gpt-4o"

# Cell 2 — create an agent
from fireflyframework_agentic.agents import FireflyAgent

agent = FireflyAgent(name="notebook-bot", model="openai:gpt-4o")
result = await agent.run("Explain quantum entanglement in two sentences.")
print(result.output)

# Cell 3 — use memory for multi-turn conversations
from fireflyframework_agentic.memory import MemoryManager

memory = MemoryManager(max_conversation_tokens=32_000)
agent_with_mem = FireflyAgent(name="chat", model="openai:gpt-4o", memory=memory)

cid = memory.new_conversation()
result = await agent_with_mem.run("My name is Alice.", conversation_id=cid)
print(result.output)

result = await agent_with_mem.run("What is my name?", conversation_id=cid)
print(result.output) # Alice

# Cell 4 — reasoning patterns
from fireflyframework_agentic.reasoning import ReActPattern

react = ReActPattern(max_steps=5)
result = await react.execute(agent, "What are the top 3 uses of Python in 2026?")
print(result.output)

# Cell 5 — structured output with validation
from pydantic import BaseModel
from fireflyframework_agentic.validation import OutputReviewer

class Summary(BaseModel):
    title: str
    bullet_points: list[str]
    confidence: float

reviewer = OutputReviewer(output_type=Summary, max_retries=2)
result = await reviewer.review(agent, "Summarize the benefits of async Python.")
result.output # displays the structured Summary object in the notebook

Tip: You do not need asyncio.run() or nest_asyncio in Jupyter — await works at the top level of any cell because Jupyter runs its own event loop.

docs/tutorial.md is an 18-chapter, hands-on guide that teaches every concept from zero to expert through a real-world Intelligent Document Processing pipeline. Start here if you want to learn the framework thoroughly.

docs/use-case-idp.md is a focused walkthrough of building a 7-phase IDP pipeline that ingests, splits, classifies, extracts, validates, assembles, and explains data from corporate documents — using agents, reasoning, document splitting, content processing, validation, explainability, and pipelines.

Detailed guides for each module:

• Architecture — Design principles and layer diagram
• Agents — Lifecycle, registry, delegation, decorators, human-in-the-loop approval
• Template Agents — Summarizer, classifier, extractor, conversational, router
• Tools — Protocol, builder, guards, composition, built-ins, native HITL approval (requires_approval, deferred resume)
• Prompts — Templates, versioning, composition, validation
• Reasoning Patterns — 6 patterns, structured outputs, output modes (output_mode/reasoning_output_mode), custom patterns
• Content — Chunking, compression, batch processing
• Memory — Conversation history, working memory, storage backends
• Validation — Rules, QoS guards, output reviewer
• Embeddings — 8 providers, auto-batching, similarity, registry
• Vector Stores — 6 backends, tenant scoping, auto-embedding, search_text, namespaces
• Pipeline — DAG orchestrator, parallel execution, checkpointing, audit log, retries
• Dynamic Workflows — Code-defined orchestration DSL over agents: @workflow, agent/parallel/pipeline/stream, budgets, journal resume, smart routing, sub-workflows, HITL, FireflyAgentRunner
• Observability — Tracing, native pydantic-ai instrumentation, metrics, events, provider-agnostic cost resolvers, budget gates
• Resilience — Circuit breaker (state machine + middleware), fast-fail on cascading failures
• Storage — Managed-SQLite durable layer: atomic writes, cross-process leasing
• Explainability — Decision recording, audit trails, reports
• Security — Prompt/output guards, at-rest encryption
• Secure Script Execution — Deny-by-default Monty sandbox, static safety pre-screen, SecureScriptRunner, Firefly Code Mode
• Experiments — A/B testing, variant comparison
• Lab — Benchmarks, datasets, evaluators
• Evaluation — LLM-as-judge metrics, RAGAS, retrieval metrics
• Studio — moved to fireflyframework-agentic-studio

git clone https://github.com/fireflyframework/fireflyframework-agentic.git
cd fireflyframework-agentic
uv sync --all-extras

uv run pytest # Run the test suite
uv run ruff check fireflyframework_agentic/ tests/ # Lint
uv run pyright fireflyframework_agentic/ # Type check

Dev dependencies (installed with uv sync): pytest >=8.3.0, pytest-asyncio >=0.24.0, pytest-cov >=6.0.0, ruff >=0.9.0, pyright >=1.1.0, httpx >=0.28.0.

See CONTRIBUTING.md for guidelines.

See CHANGELOG.md for notable changes.

Apache License 2.0. See LICENSE for the full text.