ai-agent-book/slides/lesson-39.md

theme, title, info, author, transition, mdc, lineNumbers, monaco, aspectRatio, canvasWidth, layout, class

theme	title	info	author	transition	mdc	lineNumbers	monaco	aspectRatio	canvasWidth	layout	class
seriph	Lesson 39 — When Should Agents Share the Same Context?	English video course for AI Agents in Depth	Bojie Li	slide-left	true	false	false	16/9	980	cover	cover

Expand · Chapter 10 · Multi-Agent Collaboration

When Should Agents Share the Same Context?

Shared trajectories, isolated contexts, role switching, and handoffs

Lesson 39 of 42 · 17 minutes · Classification Framework; Shared vs. Non-Shared Context; Multi-Stage Role Switching

---

Expand · Chapter 10 · Multi-Agent Collaboration

Problems this chapter will solve

Lesson 39

When Should Agents Share the Same Context?

Lesson 40

Who Should Coordinate Independent Agents?

Lesson 41

When Is Multi-Agent Actually Better Than One Agent?

Lesson 42

How Do Agent Teams Fail—and What Should We Build Next?

---

Why this problem matters

Continuity

Shared history preserves details and user decisions.

Isolation

Separate contexts reduce interference and permission leakage.

Scale

Independent contexts enable parallel work beyond one window.

---

Three ideas to keep in view

Shared context

New prompt/tools, same complete trajectory

Non-shared context

Independent trajectory + explicit communication

IPC analogy

Files as shared memory; calls/messages as message passing

---

The book's visual model

Shared-context and non-shared-context multi-Agent architectures

---

Shared context vs. Non-shared context

Shared context

• Near-zero handoff loss
• Mostly serial roles
• History grows and biases

Non-shared context

• Modular and parallel
• Selective disclosure
• Handoff can omit evidence

Share when loss is unacceptable; isolate when scale, focus, or security dominates.

---

A role is prompt + tools + visible state

stage = router.choose(task_state)
agent = Agent(prompt=stage.prompt, tools=stage.tools)
if stage.shared:
    agent.resume(full_trajectory)
else: agent.start(handoff_package)

---

Test the claim

10-11 min

Inspect staged prompts, tools, and fallback gates

Observe: Requirements, implementation, review, and revision boundaries

10-21 min

Inspect role-specific handoff capabilities

Observe: Distinct prompts, tools, transfer edges, and shared session state

Demo budget: 2 minutes · one contiguous terminal block

---

class: course-terminal

Live demo

Switching to the terminal

$ cd chapter10/staged-system-prompt && python demo.py --list-stages

$ cd chapter10/multi-role-transfer && python demo.py --list-roles

Run the command(s), narrate decisions, and point to the observation—not just the output.

---

What the evidence supports

Finding 1

Changing prompt and tools can create a specialist while preserving one trajectory.

Finding 2

Shared context removes handoff compression but can carry framing bias across roles.

Finding 3

Independent Agents require an explicit data plane and control plane.

---

layout: center

Where the claim stops

Boundary condition

Calling predefined stages 'multi-Agent' is useful architecturally, but the execution remains a workflow with a known path.

---

layout: center

Engineering takeaway

Design rule

Choose context sharing before topology: it determines information loss, isolation, parallelism, and token growth.

---

Continue the experiment

Stage-based role switching book-en/images/fig10-stage-switching.svg Staged-system full demo chapter10/staged-system-prompt/README.md Cross-domain transfer demo chapter10/multi-role-transfer/README.md

---

layout: center class: text-center

Pause and apply

Your turn

Which detail would be too dangerous to omit from a handoff—and which detail should not cross the boundary?

---

layout: center class: text-center

Next · Lesson 40

Choose who coordinates independent Agents and how artifacts and messages move between them.

→