Patterns

A single HazelJS AgentRuntime call enriched with retrieval, bounded tools, or structured output—without an open-ended agent loop.

A bounded AgentRuntime observe→decide→act loop where HazelJS validates tool calls via ToolRegistry and feeds observations back to the model.

Pause a HazelJS AgentRun with durableSuspend before high-impact tools until a human approves or rejects via approveAndResume.

Iterate generation with HazelJS AgentRuntime loop stages (plan/execute/critique/validate) until quality bars or budgets are met.

Persist AgentRun checkpoints and leases with createDurableRunStore / createSqlDurableRunStore so HazelJS agents survive crashes without double-applying side effects.

Separate working context, session state, long-term memory, and artifacts in HazelJS so agents scale without stuffing everything into one prompt.

Make AgentRuntime write tools safe under retries and durable resume using idempotency keys and dedupe semantics in adapters.

Undo or neutralize partial side effects after agent or workflow failure using saga/flow compensation and idempotent reverse tools.

Cap what a HazelJS agent may do unsupervised using capabilities, PolicyEngine effects, maxSteps, and HITL—not prompt instructions alone.

Enforce hard economic stops with RunBudget maxTokens/maxCostUsd plus maxSteps—budgets that live in AgentRuntime, not in prompt text.

Reduce blast radius for HazelJS agents using session/tenant isolation, curated tools, and policy—without claiming an unreleased OS/container sandbox product.

Route intents to specialized agents or subgraphs with AgentGraph conditional edges or SupervisorAgent.

Fixed-stage pipelines with @hazeljs/flow nodes or runtime.pipeline—when stages are stable.

Run independent branches concurrently with AgentGraph parallel nodes, then merge.

Produce an explicit plan artifact, then execute under validation—via loop stages or split agents.

Canonical observe→decide→act loop owned by AgentRuntime—not ad-hoc while(true) in app code.

Require a structured critique pass before committing answers or side effects—via loop stages or evaluator–optimizer.

A supervisor delegates to specialist agents with narrow tools via createSupervisor / AgentGraph.

Transfer goals and minimal state between agents or to humans without dumping full transcripts unsafely.

Require grounded lookup/RAG before write tools fire—via policy, graph edges, or tool allow-order discipline.

Choose models by task class, cost, and risk using per-agent model prefs and documented cost routing—not one frontier model for everything.

Reference architecture for a HazelJS production Agent Runtime: App code, DNA, Store, platform desired state, and AgentRuntime execute — as taught by Meridian Ops.