Flyto2 Blueprint

Most agent memory stores what was said. Blueprint stores what actually worked.

Turn AI work into verified, replayable procedures.

AI said it finished. Flyto2 shows the proof.

Product contract: flyto.product-contract.v1 · Product: Flyto2

Flyto2 is one product delivered as three independently usable packages:

flyto-blueprint is an independently usable procedure-memory package. Language and LLM systems may understand or route work; Blueprint persists, expands, and ranks procedures from validated execution outcomes supplied by a host. This evidence-first boundary means Blueprint neither executes workflows nor guesses scientific facts. Mathematics, physics, chemistry, robotics, and other domain solvers belong in bounded capabilities and hosts when evidence requires them.

Flyto2 Blueprint is procedure memory for AI agents. It turns a successful execution into a parameterized workflow that can be searched, run again, and judged by its real history.

successful execution
        ↓
parameterize what changes
        ↓
save steps + retries + assertions + compatibility
        ↓
reuse with new arguments
        ↓
record the outcome in an Evidence Card

It is closer to turning a good agent run into a tested function than adding another chat-history or vector-memory layer.

Blueprint does not train model weights. It makes a verified procedure reusable by an execution host.

Installation and first run

pip install flyto-blueprint

from flyto_blueprint import BlueprintEngine, MemoryBackend

engine = BlueprintEngine(storage=MemoryBackend())
result = engine.expand("browser_scrape", {
    "url": "https://example.com",
    "extract_selector": "h1",
})
print(result["data"]["steps"])

How this differs from a typical AI agent

An individual Evidence Card stays deliberately narrow: planner_model_calls_used=0 means exact reuse skipped outer-agent planning; it does not pretend an llm.* workflow step was free. The v3 benchmark goes further by recording native planner and workflow counters separately, checking their arithmetic, and making a full-usage claim only when both are observed.

Proof, not vibes

Every Blueprint summary includes an Evidence Card. It shows the number of trusted outcomes, observed success rate, Wilson 95% lower bound, retries, assertion pass rate, p50/p95 duration, and measured zero-planner-call reuse. Detailed samples keep only an allowlist of execution facts and are capped to the latest 100 entries; prompts, parameters, API keys, and raw results are not accepted.

Search and list summaries also expose ordered, unique module_ids. This field contains capability names only—never step parameters or results—so Flyto2 AI can use trusted prior workflows as routing hints without expanding or executing them.

An Evidence Card answers “did this procedure work?” The versioned effectiveness benchmark answers the harder question: “does Blueprint help an agent without making it less reliable?” It compares the same tasks, model, environment, and random seeds across four modes, including ordinary conversation, Traditional Chinese and Japanese negation, hostile evidence, incompatible reuse, and a sealed holdout.

The published v3 result is 10 tasks × 20 trials × four paired modes across five runs: 4,000 raw records, three model families, Apple Silicon and Linux x86-64, and one independent GitHub runner. The workloads perform real Python file/test execution, real loopback HTTP browser and API I/O, real filesystem persistence, and real Ollama inference. No planner or workflow call is mocked.

Across all five runs, warm verified reuse kept benchmark and workload success at 100%, with zero manual corrections and zero false reuse. Compared with Flyto2 routing without Blueprint, total observed model tokens fell by 71.25–72.90%. Compared with the generic agent baseline, they fell by 84.80–85.78%. The paired 95% lower bound for the Blueprint-specific reduction was 63.29–64.43%, so the result does not depend on the point estimate alone.

That is evidence for this controlled suite, these model bytes, and these hosts; it is not a claim that every task gets 70% cheaper. Read the one-screen result story, inspect the committed raw JSONL, or rerun the deterministic closure verifier.

Official links: flyto2.com · Docs · PyPI · flyto-core · flyto-ai

Good fit if you searched for:

• reusable AI workflow patterns
• workflow automation blueprint engine
• self-learning automation recipes
• YAML workflow templates for AI agents

What it already does

• 33 built-in browser, API, data, image, notification, monitoring, PDF, and OCR patterns.
• Synonym-expanded search, so “grab” can find “scrape.”
• Repository/runtime compatibility, so similar-looking workflows do not get mixed across incompatible projects.
• Retry and assertion contracts that survive learning and expansion.
• Evidence Cards that expose reliability and zero-planner-call reuse.

Learn, measure, and share

from flyto_blueprint import BlueprintEngine, MemoryBackend

engine = BlueprintEngine(storage=MemoryBackend())

# List available blueprints
blueprints = engine.list_blueprints()

# Expand a blueprint with arguments
result = engine.expand("browser_scrape", {
    "url": "https://example.com",
    "extract_selector": "#content",
})

# Learn from a successful workflow
engine.learn_from_workflow(workflow_dict, name="My Pattern", tags=["browser"])

# Report a trusted runtime outcome with measured facts
engine.report_outcome(
    "my_pattern",
    success=True,
    execution_id="run-123",
    evidence={
        "duration_ms": 842,
        "step_count": 3,
        "total_attempts": 3,
        "assertion_passed": True,
        "selection_mode": "deterministic",
        "planner_model_calls_used": 0,
        "model_call_scope": "planner",
    },
)

# Share explicitly; the library never uploads on its own
bundle = engine.export_blueprint("my_pattern", publisher="my-team")
engine.import_blueprint(bundle["data"])

Unsigned or unknown-publisher imports are quarantined as community. A host may sign exports and configure trusted publisher keys through the Python API; signing keys are intentionally unavailable to model-facing tools.

Usage

Use Flyto2 Blueprint when an AI agent should reuse a known workflow shape instead of generating a brand-new sequence every time. Typical use cases:

• Browser scrape, screenshot, and form-fill recipes.
• API integration workflows with typed arguments.
• PDF, OCR, image manipulation, and notification patterns.
• Learned workflow reuse for teams that run similar automations repeatedly.

API

The package facade, engine lifecycle, storage protocol, scoring behavior, and complete declaration inventory are documented in API and the generated Python reference. MCP consumers should use the generated tool reference. Benchmark hosts should start with the plain-language benchmark guide.

Architecture

Architecture explains discovery, expansion, persistence, learning, and trust boundaries. Features maps each behavior to its implementation and tests; the whitepaper explains the design rationale and limits.

Configuration

The library has no mandatory remote service. Storage-specific credentials and runtime settings belong to the selected backend and deployment secret store; never place them in blueprint YAML or committed examples.

Storage Backends

BlueprintEngine() takes no backend by default, and with no backend nothing is persisted: learned blueprints and reported outcomes live for the lifetime of the process and are then discarded. Pass one explicitly to keep them.

• MemoryBackend — In-memory, great for tests
• SQLiteBackend — File-based persistence; safe across threads and processes (WAL plus an immediate transaction around every read-modify-write), so several hosts may share one database file
• FirestoreBackend — Google Firestore (for flyto-cloud)

Testing

python -m pytest
python -m ruff check .
python scripts/benchmark-scorecard.py verify-results \
  --suite benchmarks/suites/blueprint-effectiveness-v1.yaml \
  --results-dir benchmarks/results
python scripts/benchmark-scorecard.py verify-results \
  --suite benchmarks/suites/blueprint-effectiveness-v2.yaml \
  --results-dir benchmarks/results/blueprint-effectiveness-v2
python scripts/benchmark-scorecard.py verify-results \
  --suite benchmarks/suites/blueprint-effectiveness-v3.yaml \
  --results-dir benchmarks/results/blueprint-effectiveness-v3
python scripts/run_longitudinal_evidence.py verify \
  --evidence benchmarks/results/longitudinal/local-longitudinal.evidence.json

Contributing

Open an issue or pull request for new blueprint categories, scoring behavior, storage backends, docs, or examples. Security reports should go to security@flyto2.com.

License

Apache-2.0