Effect Agent

Appearance

● All roadmap phases (P0–P7) implemented · pre-1.0, specifications Draft

Agents,with cause.

Build autonomous TypeScript agents without leaving Effect behind. Schemas stay canonical, failures stay typed, dependencies stay visible, and every resource stays inside a Scope.

See one complete Agent →Build it step by step

The whole runtime contract

Effect<A, E, R>
A
schema-decoded output
E
AI + tool + domain failures
R
model + handlers + services

No hidden runtime. The signature is the architecture.

AvailableBounded interpreter

Run, stream, tools, policy, and explicit model binding.

AvailableOperational capabilities

Safe-seam input, approval, budgets, context, MCP, sandbox.

AvailablePersistent Conversations

Replay, checkpoints, export, memory and SQLite stores.

AvailableDurable accepted work

Receipts, Attempts, fencing, recovery, and Settlement on Node/SQLite and Cloudflare.

Two ways to read these docs

Use what exists. Pressure-test what comes next.

Current guides are backed by code and phase evidence. Future guides present the intended public interfaces before they exist, so the architecture can be judged from the outside. Every page carries its implementation status.

01 / ORIENT

See the whole Agent

Read one end-to-end file from Schema and Tool definition through Layers and execution.

Open the example →02 / EXECUTE

Run or observe

Use one interpreter as an Effect result, semantic Stream, or scoped detached Run.

See runtime APIs →03 / REMEMBER

Persist Conversations

Append canonical records, rebuild projections, and resume observation without claiming durability.

Explore persistence →04 / RECOVER

Design for interruption

Follow the implemented path from durable admission to one honest terminal Settlement.

Inspect the durable runtime →05 / PROVE

Test without a provider

Script model Turns, inject Layers, control time and IDs, and verify the semantic trace offline.

Open the test kit →06 / VERIFY

Check every claim

See the package, phase, evidence, maturity label, and explicit non-claim behind each surface.

View status →

Architectural axioms

The framework does less so Effect can keep doing more.

Effect Agent owns the multi-Turn loop, Conversation, and recovery model. It deliberately does not duplicate Effect AI, invent another dependency container, or turn external side effects into magical exactly-once claims.

01 / SCHEMA

One source of truth

Input, output, Tool, wire, and stored values begin as Effect Schemas. Provider shapes are derived.

02 / LAYER

Requirements stay visible

Models, Tool handlers, stores, clocks, sandboxes, and policies arrive through explicit Layers.

03 / SCOPE

Nothing outlives its owner

Model streams, Tool fibers, queues, clients, and attached children belong to the Run Scope.

04 / RECORD

Recovery stays honest

Replay rebuilds state, never external effects. Ambiguous outcomes remain explicit rather than guessed.