PrerequisiteAI 202 recommended, plus Modules 1 to 4 of this course
0 / 6 completed
Sync my progressJoin a session to keep progress across devices and share it with your instructor
Learning objectives
Explain the difference between Shared Context and Private Context, and why a team of agents needs both
Describe Shared Memory, Local Memory, Shared Knowledge and a Source of Truth as concepts
Set Information Boundaries using ownership, access control and the five information classes
Recognize stale shared context and describe context freshness, version awareness and synchronization
Explain Coordination State and consensus concepts, then complete Activity 5 and produce a Shared Context Map
Shared and private context
9 min
Share everything, or share only what is needed?
In Module 4 agents learned to send each other messages. Now a harder question appears: what information should the agents hold in common? A team that shares nothing cannot work together. A team that shares everything wastes effort, leaks information and drowns each agent in noise.
Core question
Should multiple agents share all information, or should each agent see only the context it needs?
Key terms
Term
Plain meaning
Shared Context
Information that every agent on the team is meant to see, such as the goal and agreed facts
Private Context
Information that belongs to one agent because only its own role needs it
Information Boundary
A deliberate line that says which information may cross between agents, or to a human
Context ownership
The one role responsible for keeping a piece of information correct
Simulated example
A simulated customer-feedback team has a Research agent, an Analysis agent and a Reporting agent. All three need the shared goal and the list of approved sources. Only the Analysis agent needs its raw working notes. Only a human reviewer should see a draft that mentions an unresolved complaint.
Info
This module stays conceptual. We design what is shared and what is private. We do not build storage, checkpoints or recovery systems, which belong to AI 302.
Key takeaway: Multi-agent teams need a deliberate answer to what is shared and what is private, not a default of sharing everything.
9 min
Context architecture: shared, private and bounded
A context architecture starts from a Shared Goal. Below it sits one Shared Context that acts as the Source of Truth. Each agent then keeps its own Private Context beside it.
Context architecture
Shared Goal: the purpose every agent works toward
Shared Context / Source of Truth: agreed facts, task status and approved decisions
Agent A Private Context: only what Agent A needs for its role
Agent B Private Context: only what Agent B needs for its role
Agent C Private Context: only what Agent C needs for its role
Info
Text version of the diagram: the Shared Goal is at the top. Under it is one Shared Context that serves as the Source of Truth. From that Shared Context, three branches lead to Agent A, Agent B and Agent C. Each of the three agents has its own Private Context that the other agents do not see by default.
Five classes of information
Class
Meaning
Example (simulated)
Shared context
Every agent on the team may see and rely on it
The goal, the approved source list, the report outline
Private agent context
Visible to one agent only, because only its role needs it
The Analysis agent's scratch calculations
Shared knowledge source
A reference many agents read but none owns as working state
A policy handbook or product FAQ
Human-only
Reserved for a person to see and decide on
A draft that names an unresolved complaint
Restricted
Not to be passed to any agent without explicit approval
Personal data, credentials, confidential figures
Check yourself
An agent's rough working notes are only useful for its own step. Where do they belong?
Key takeaway: Share the goal and agreed facts, keep role-specific working material private, and classify each piece of information.
Memory, freshness and synchronization
9 min
Shared Memory, Source of Truth and ownership
Memory in a multi-agent team comes in two conceptual kinds. Shared Memory holds what the team has agreed or produced together. Local Memory holds what one agent wants to remember for its own work. Shared Knowledge is different again: it is reference material many agents read.
Memory and knowledge concepts
Concept
Who uses it
Design question
Shared Memory
The whole team
What has the team agreed on, and who may change it?
Local Memory
One agent
What does this agent need to remember that others do not?
Shared Knowledge
Many agents, read only
Which reference sources are approved for everyone?
Source of Truth
The whole team
When two agents disagree about a fact, which record wins?
Ownership and access control
Question
Design rule
Who owns this piece of information?
Exactly one role owns it and keeps it correct
Who may read it?
Only roles that need it for their task
Who may write or change it?
Usually the owner, or a role the owner has approved
Who reviews changes to important facts?
A named role, and a human for high-impact facts
Warning
Two owners means no owner. If two agents can both rewrite the same fact, they will eventually contradict each other. Give each shared fact one owner and one Source of Truth.
Check yourself
Why should each shared fact have one owner?
Key takeaway: Distinguish shared memory, local memory and shared knowledge, and give every shared fact one owner and one Source of Truth.
9 min
Stale context, freshness and synchronization
Shared context changes over time. When one agent updates a fact and another keeps using the old version, the team works from stale shared context. The result can look confident and still be wrong.
Simulated example
The Research agent updates the approved source list to remove a source that turned out to be outdated. The Reporting agent, still holding the earlier list, cites that source in its draft. Nobody made a mistake alone; the two agents were not synchronized.
Ways to reduce stale context (concepts only)
Idea
Plain meaning
Context freshness
How recent a piece of shared information is, and how long it stays trustworthy
Version awareness
Each shared fact carries a version or time so an agent can tell if it is holding an older copy
Context synchronization
A planned moment when agents refresh their view of the shared context
Change notice
A message, using the Module 4 contract, that tells others a shared fact changed
Info
Text version of the idea: an agent that reads shared context should know how old it is. An agent that changes shared context should tell the others. A refresh point before an important step lets agents compare their versions with the Source of Truth.
Warning
Boundary: this is about designing for freshness. Building storage, checkpoint or recovery machinery to guarantee it is a later topic in AI 302.
Check yourself
Which practice best helps an agent notice it is using an outdated fact?
Key takeaway: Attach version or time to shared facts, tell others when they change, and refresh before important steps.
Coordination, practice and deliverable
10 min
Coordination state and consensus concepts
Besides facts, a team needs to know where the work stands. Coordination State is the shared picture of who is doing what, which steps are done, and which are waiting. It is a design idea here, not a database schema.
What a Coordination State might show (conceptual)
Item
Question it answers
Task status
Which tasks are open, in progress or finished?
Assignment
Which role holds each task now?
Waiting on
Which task or human decision is blocking progress?
Decisions made
Which choices has the team settled, and by whom?
Open questions
Which points remain undecided?
Sometimes agents hold different views and the team needs one answer. Consensus, as a concept, is a rule for reaching a shared decision. Rules can be simple: the owner decides, the supervisor decides, a majority of reviewers agree, or a human decides.
Consensus concepts compared
Approach
Strength
Watch out for
Owner decides
Fast and clear accountability
One agent's blind spot becomes the team's
Majority of reviewers
Several checks on one answer
Agents built the same way may repeat the same mistake
Human decides
Judgment and accountability stay with a person
Slower, so reserve it for high-impact points
Warning
Agreement among agents is not proof of correctness. For high-impact decisions, a human still reviews and remains responsible.
Check yourself
Three agents agree on an answer that will affect a customer contract. What is the safest conclusion?
Key takeaway: Coordination State shows where work stands, consensus is a rule for one answer, and humans keep high-impact decisions.
9 min
Activity 5: Shared or Private? and the Shared Context Map
Take a simulated multi-agent team, or the team you designed in Modules 2 to 4. Sort its information into the five classes and decide who owns and who may see each item.
1List the information Write 8 to 12 pieces of information the team uses or produces.
2Classify each item Choose one: shared context, private agent context, shared knowledge source, human-only or restricted.
3Assign an owner Name the one role responsible for keeping each shared item correct.
4Set access and boundaries Note which roles may read it and which may change it.
5Plan freshness For each shared item, note how agents will know it changed and when they refresh.
Warning
Privacy reminder: use simulated or already-public information. Do not enter company secrets, client personal data or unauthorized sensitive information.
Info
Deliverable: a Shared Context Map. It lists each information item with its class, owner, who may read, who may change, freshness plan, and the coordination state the team tracks. Keep it for your Multi-Agent Architecture Canvas.
