Context and memory

The next agent knows the context you give it, plus what its tools and environment make available. A handoff between agents is an explicit data-flow decision.

Choose what crosses the handoff

Raw material → researcher → typed brief → writer → draft
                               ↑
                    selected facts + uncertainty

Passing a structured brief is often easier to inspect than forwarding an entire transcript. Keep sources and uncertainty attached to the relevant claims.

import { Schema } from "effect";

export const Brief = Schema.Struct({
  question: Schema.String,
  facts: Schema.Array(Schema.Struct({ claim: Schema.String, source: Schema.String })),
  uncertainty: Schema.Array(Schema.String),
});
export function writingPrompt(brief: typeof Brief.Type): string {
  return [
    "Write a short draft from this brief.",
    "Do not invent supporting evidence. Preserve material uncertainty.",
    JSON.stringify(brief),
  ].join("\n\n");
}

Pass writingPrompt(brief) to the writer's session. This function only assembles context; it does not retrieve documents or store memory. Treat document contents as evidence to assess, not instructions that can change the workflow's policy.

Conversation, files and memory

There is no native cross-run memory API in published alpha.13. A long conversation is also not an unlimited context window. For long tasks, return typed progress such as completed tasks, pending tasks and the current plan, then choose what the next step receives.

If retrieval can change over time, record its result at an activity boundary when the run needs a reproducible snapshot. Keep a source revision or capture date with it. A URL alone does not preserve the page you read.

Skills are not a filesystem boundary

Docker supports selected injected skills. Local execution uses native installed skills and rejects injection. Neither “selected skills” nor a prompt proves that other instructions or tools are inaccessible.

Reference: Agent settings · Session prompts

← Execution and isolation · Next: Studio, logs and usage →