CS2680 pattern catalog
Eight small Community blocks, one per multi-agent architecture from the CS2680 Agent Design II lecture.
The pattern catalog is a set of eight small teams. Each one shows one way to arrange agents, on a fictional task small enough to finish in minutes. They are teaching exercises: run one, watch how the agents coordinate on the board, then reuse the shape for real work.
The architectures come from Juncheng Yang's Agent Design II slides. The Jobs are original Sculptured instructions, not course code, and the course does not endorse them.
Find the blocks
Open Community → Explore and search for cs2680. Every block in the catalog carries that tag. Inspect a block to read its complete Jobs, then Save block or Add to project. See Community for the full steps.
The eight blocks
| Block | Slide | Agents | What to try |
|---|---|---|---|
| Supervisor & workers | 9 | Supervisor, Benefit analyst, Feasibility analyst | One supervisor delegates two independent analyses of a fictional park proposal and integrates their evidence. |
| Hierarchical feature team | 10 | Feature manager, two leads, four specialists | A manager coordinates a backend lead and a frontend lead; each lead scopes two specialists. They design a tiny reading-list feature without building it. |
| Pipeline with feedback | 11 | Field extractor, Arithmetic checker, Summary writer | Extract a fictional invoice, check the arithmetic, then summarize. One bounded retry sends a concrete correction upstream. |
| Shared workspace blackboard | 12 | Feature author, Risk author, Check author | Three peers coordinate through separate versioned files in one folder. Each file has one writer, so nobody overwrites another's work. |
| Independent review committee | 13 | Two reviewers and an editor | Two reviewers assess the same short argument without seeing each other's verdicts; the editor compares the evidence. |
| Capability bidding exercise | 14 | Auctioneer and three bidders | Fictional workers bid in toy cost and time units; the auctioneer checks their constraints and awards one task. |
| Peer negotiation swarm | 15 | Three pickers | Three peers agree who owns which fictional picking task, with no permanent coordinator. |
| Dynamic specialist dispatch | 16 | A coordinator, then up to two children | The coordinator creates billing or log specialists only when the fictional support ticket calls for them. |
flowchart LR
subgraph sup["Supervisor & workers"]
S["Supervisor"] --> W1["Analyst"]
S --> W2["Analyst"]
W1 --> S
W2 --> S
end
subgraph pipe["Pipeline with feedback"]
E["Extractor"] --> C["Checker"] --> Su["Writer"]
C -.->|"one correction"| E
end
subgraph com["Review committee"]
R1["Reviewer"] --> Ed["Editor"]
R2["Reviewer"] --> Ed
end
Before you run one
- Choose a folderEvery block is portable: it has no folder, no model choice and no saved session. Set a project folder before starting.
- Read the JobsEach agent's Job says what it may and may not do. Some, such as the blackboard's authors, write files in the project folder.
- Mind the costThe agents are Codex agents. Starting them uses your own Codex plan. Browsing and saving blocks costs nothing.
What the arrows mean here
- Hands work to arrows use Sculptured's scheduler: the giver starts first and hands its work on.
- Reports to and watches arrows record design intent. The agents coordinate through their Jobs, using managed messages and files.
- A planned arrow does not prove that a message was sent. Open the team's communications to see what agents actually sent each other.
The dynamic dispatch block contains only the coordinator. Its specialists appear on the Run board only when the coordinator actually creates them with spawn_agent.
From exercise to real work
Once a shape fits your problem, adapt it:
- Save the block to your Library and add it to a project.
- Rewrite each Job for your task, using the Job checklist.
- Add a check, or end the team with an agent that runs the tests, so "done" is checked rather than only reported.
- Set token limits before you add more agents.