Sylo SDK

“When autonomous AI agents move from local prototypes to production, unexpected edge cases arise: rate limits trigger crashes, hallucinations touch unauthorized data, and destructive actions fire without oversight.”

01 / RESTART FAILURE

Pipelines crash and burn tokens

A 5-step workflow crashes at step 4 due to an API timeout. Without Sylo, you restart from step 1, burning duplicated LLM tokens and wasting minutes of latency.

02 / UNBOUNDED ACCESS

Agents over-privilege and leak

An autonomous assistant hallucinates and accesses internal customer databases or AWS S3 buckets it shouldn't touch. You find out after the damage is done.

03 / NO OVERSIGHT

Irreversible actions fire silently

Deleting records, initiating financial wire transfers, or modifying production schemas occur automatically without any notice or mechanism for human sign-off.

HOW IT WORKS

Three lines of code. Four production guarantees.

pipeline.py

import sylo

sylo.init(project="customer-operations", environment="production", storage="redis")

@sylo.step("fetch-customer", max_retries=3)
@sylo.trust(can_read=["crm.customers"])
async def fetch_customer(ctx, customer_id: str):
    return await ctx.access("crm.customers", action="read", handler=crm.get)

@sylo.step("delete-account")
@sylo.requires_approval(action_class="destructive", notify=["slack", "email"])
async def delete_account(ctx):
    return {"status": "deleted", "id": ctx.metadata["customer_id"]}

async with sylo.pipeline("churn-remediation") as pipe:
    await fetch_customer(pipe.context, "cust_8842")
    await delete_account(pipe.context)

✓ Checkpoint Recovery

✓ Zero-Trust Enforcement

✓ Human Approval Gates

✓ Immutable Audit Log

CORE FEATURES

Engineered for mission-critical reliability.

01 / RECOVERY & COST LEDGER

Smart Checkpointing & Transparent Cost Modeling

Every completed step is serialized to storage. If downstream API limits trigger a crash, reruns instantly return cached results. Built-in pricing tables calculate exact token counts and estimated USD savings ($2.50/$10 per 1M tokens for gpt-4o) without requiring live API keys.

ctx.record_token_usage(prompt_tokens=1200, completion_tokens=350, model="claude-3-5-sonnet")

02 / ZERO-TRUST BROKER

Granular Capability Manifests

Prevent hallucinating agents from executing unauthorized network requests or file deletions. Declare explicit scopes per step (`can_read`, `can_write`, `can_delete`). Any out-of-bounds access raises a SyloPermissionError and logs an audit violation instantly.

@sylo.trust(can_read=["slack.channels"], can_delete=[])

03 / APPROVAL GATES

Asynchronous Human Oversight & Local Dev Server

Guard financial transfers or schema deletions behind explicit human sign-off. In local development, Sylo spins up a clean background HTTP server on port 7749, generating clickable approval URLs right in your terminal before executing dangerous code.

Approve: http://localhost:7749/approve/d3b07384...

04 / AUDIT TRAILS

Immutable Event Ledger via Redis Streams

Every pipeline execution produces an append-only sequence of events (`PIPELINE_STARTED`, `STEP_COMPLETED`, `APPROVAL_DECISION`). Persisted locally as clean JSONL or streamed to high-throughput Redis Streams (`XRANGE`) for SOC2 enterprise compliance.

~/.sylo/executions/{id}/audit.jsonl

STORAGE BACKENDS & CONFIGURATION

Standardized interfaces from dev to production.

Backend Driver	Configuration	Use Case & Persistence Guarantees
Local File Store	storage="local"	Default engine. Perfect for local development and CI testing. Stores execution manifests, checkpoints, and JSONL audit logs under `~/.sylo/executions/`. Fails silently on network errors so developer workflows are never blocked.
Redis & Streams	storage="redis"	Built for high-concurrency production workloads. Uses Redis key-value storage for sub-millisecond state checkpoints and append-only Redis Streams for real-time audit event distribution.
Sylo Cloud API	storage="cloud"	Enterprise managed platform. Syncs executions, approval gate notification queues, and real-time telemetry directly to your centralized team dashboard.

ENVIRONMENT VARIABLES

Variable	Default Value	Description
SYLO_PROJECT	"default"	Identifier grouping related pipelines and agent workflows.
SYLO_ENVIRONMENT	"development"	Runtime target (`development`, `staging`, `production`).
SYLO_STORAGE	"local"	Storage driver persistence selection (`local`, `redis`, `cloud`).
SYLO_API_KEY	None	API key when communicating with Sylo Cloud.
SYLO_REDIS_URL	"redis://localhost:6379"	Connection string when using Redis backend driver.

Verified against real AI infrastructure

Sylo wraps your existing async nodes and agent loops cleanly without replacing your orchestration architecture. Every adapter is 100% E2E verified against live inference infrastructure.

LangGraph · SyloGraph OpenAI Agents · wrap_agent CrewAI · SyloCrew Vanilla Python

Subsystem	What Was Tested	Status
Checkpoint Recovery	Real Groq LLM call → crash → resume → completed step skipped from disk checkpoint	✅ Verified
Approval Gates	Real HTTP server on localhost:7749 → auto-approve click → pipeline resumed	✅ Verified
Framework Integrations	Live Groq inference verified across LangGraph StateGraph, OpenAI Agent Runner, and CrewAI Crew/Task	✅ Verified
Trust Broker & Audit Engine	ctx.access() runtime enforcement, get_summary(), replay(), and immutable JSONL / Redis Stream audits	✅ Verified

EARLY ACCESS

Be first when Sylo Cloud launches.

Get three months of Pro free, direct access to the founder, and influence over roadmap prioritization.

✓ You're on the list.

We'll be in touch when Sylo Cloud opens.