Rewyn documentation¶

Rewyn is a provider-agnostic SDK for building AI agents and, more importantly, for understanding what they did afterwards. These pages teach the engineering, not just the call signatures.

Every page follows the same shape: what the thing is and why it exists, the smallest useful example, what it looks like in production, where to find the API, and the ways it goes wrong.

Start here¶

Getting started

Building¶

Page	What it covers
Models	One interface across providers, cost and structured output
Agents	The loop that turns a model plus tools into behaviour
Tools	Typed, versioned, permissioned functions a model can call
Context engineering	Deciding what goes in the prompt, and proving it later
Memory	What the system remembers between runs
RAG	Retrieval as an observable pipeline, not a black box
MCP	Connecting to Model Context Protocol servers
Skills	Versioned instructions loaded from `SKILL.md`

Controlling¶

Page	What it covers
Loops	ReAct, plan-execute, reflection, and budgets
Graphs	Explicit control flow when a loop is not enough
Subagents	Delegating a scoped task
Handoffs	Transferring a conversation to another agent
State	Versioned, snapshot-able execution state
Sandbox	Running generated code without trusting it
Guardrails	Input and output policy, recorded as decisions
Human in the loop	Approvals, feedback and corrections

Improving¶

Page	What it covers
Replay	Reproducing a run exactly, or changing one component
Evaluation	Deterministic metrics, LLM judges, custom evaluators
Regression	Datasets, baselines and release gates

Operating¶

Page	What it covers
Console	The local UI and the cloud console: reading what a run did
Production	Recording, sampling, cost, migrations and reliability
Integrations	Redis, S3, search, vector databases, other frameworks
Security	Secrets, trust boundaries and prompt injection
Cloud	The optional hosted platform

The whole thing at once¶

examples/enterprise_research_agent/ is the golden demo: one application that uses every subsystem above and then replays, diffs, evaluates and gates itself. It runs offline.

uv run python -m examples.enterprise_research_agent