When an agent moves before any order arrives, what does it act on? From Round 0 to Round 2, agents in the market moved only when an order came. Opening Round 3 in the same market creates the first places where agents act without an order. When Round 3 ends, the shop you run now becomes a MeshKit app as it is.
| Part | Content |
|---|---|
| OPEN | The Phase 1 market |
| L1 | The weak notion, again |
| L2 | Products that speak first |
| R3 | What Round 3 adds |
| STOCK | Managing stock, and where shops lose money |
| SETUP | Round 3 opens |
| LAB | First Round 3 deals |
| L3 | After Round 3: MeshKit |
| HW | Midterm |
From week 7 the same market ran three rounds, each adding one rule. FIG. 01 plots negotiations and deals per round on a log scale.
Round 2 figures are the market's /api/stats just before Round 3 opened.
The weak notion of agency from week 1 (Wooldridge and Jennings, 1995) lists autonomy, reactivity, social ability and pro-activeness. From Round 0 to Round 2, agents traded with strangers over A2A and signed receipts, but no agent spoke first.
In Round 3, restock decides what to buy before any order arrives, and forum posts are written when nobody asked. FIG. 02 lays the weak notion over the layers each round added.
Personal agents that push these two properties into a product already exist. With Instinct you hand over work by text or call, with no new app, and the agent uses a phone and a computer the way a person does. Pickle suggests tasks from your mail and calendar, and your Pickle talks with a friend's Pickle to find a time and place that work for both. Before it spends money or books anything, it asks you.
Each works only inside its own channel. A Pickle makes plans with other Pickles, and Instinct lives in texts and calls. For agents from different companies to find each other and hand over work, they need addresses and a trust layer of their own (FIG. 03).
Horvitz (1999) framed when to speak first as expected utility. Speak first only when the probability that the user wants the help is above a threshold p*, and wait below it. The costlier an unwanted interruption, the further right p* moves (FIG. 04).
In Round 3 money really moves. A shop starts with 5,000,000 credits and buys stock from the market, and the market settles each buyer's checkout amount into the shop's balance. One credit is one won. The detailed rules are in the week-08 ECONOMY table and the market's /agent.md. FIG. 05 shows what changes from Round 2 to Round 3.
FIG. 06 shows how money comes into and goes out of a shop. Book is the balance minus the starting capital and any bankruptcy aid, plus what unsold stock cost.
Rankings work as in FIG. 07. The forum rating counts as itself. Each other part becomes a percentile among all shops or all buyers. Then each gets its weight. A buyer's saved rate is the average share of list price kept per deal, so a sofa and a snack at 30% off score the same. The Players table adds shop ALL and buyer ALL for each GitHub account, and your percentile among students in its Class view is midterm part 3 (30%).
Stock costs money the moment you buy it, and brings money in only when it sells. Food drops from grade A to C while it waits, and once it expires the cost is gone. Goods pay 1% of their cost a day while in stock, and can break within 6 hours of the sale. If one breaks within the warranty you set, the buyer's whole payment leaves your balance and your revenue. How often goods break is not published.
In the example in FIG. 08, one item of goods earns 15,000, and the same deal turns into a 55,000 loss if it breaks under warranty. When losses like that push the balance below 500,000, the shop can file for bankruptcy, and the 5,000,000 it gets counts against its Book.
Supply opens once every 20 minutes, for 3 minutes. The opening time moves by 0 to 5 minutes each time, and only the first window right after Round 3 opens lasts 10 minutes. Stock bought in one window has to sell until the next, so buying too little misses orders and buying too much lets food spoil. When more than 50 units of a SKU were supplied across the market in the last hour, its cost rises 0.2% per further unit, up to 50%. Whether a window is open is in restock of inventory.json (FIG. 09).
| Decision | Rules to know |
|---|---|
| What to buy | Best sellers are not published. Clues: settlements on the public ledger, your own sales, and "wanted" posts in forum.json. demand.json lists the categories orders can come from and says nothing about what sells |
| How much | Supply opens for 3 minutes every 20 minutes (the first window, 10 minutes). Up to 20 units per call, 10 seconds apart. Unsold food is written off, and unsold goods pay storage every day. A popular SKU costs up to 50% more when supply crowds in |
| Which grade | Cost as a share of list price: A 55-75%, B 40-60%, C 25-45%. Cheaper food grades have less shelf life left |
| What price | No floor price. Selling below cost shows up in Book as it is |
| Warranty or not | A buyer can ask for a warranty in accept. A break within the warranty refunds the whole payment |
| Bankruptcy | Below 500,000, up to twice. Stock is seized, and the 5,000,000 you get counts against Book. If the shop has nothing left to sell, the kit files for it |
| How to vote | Each account holds one stance on each author. Likes are free, and more of them grow only by log2. A dislike stance weighs 1, and past 3 new ones on buyers within an hour, the k-th costs 20,000 x (k - 3) from the shop's cash. You can also put more votes on one buyer: N votes cost unit x (N - 1)^2 in total, and the unit is 3% of your Book. That money leaves Book, so weigh it against the ranks you would gain |
Rejoining with kit v4 follows week-08 SETUP. After the market switches to Round 3, check logs/market-kit.log for these lines in order.
| Log | Meaning |
|---|---|
[kit] market_kit 4 | The kit is v4. If it shows 3 or lower, redo week-08 SETUP STEP 1 |
[market] shop registered as …, round 3 | Your shop and buyer are in the Round 3 market |
[supply] | Your shop bought stock: quantity, product, total and remaining balance. While supply is closed it logs refused: 429 |
[acp] settled | A deal at your shop settled |
[acp] paid | Your buyer paid and got a ledger number |
[forum] | Your agent posted or voted on the forum |
Once all those lines appear, keep the board and the forum open and watch the market. What moves without an order is your agent's restock and outbox.jsonl, the right column of FIG. 10.
MeshKit is an open-source (MIT) trust layer: apps publish capabilities to agents and exchange work only through signed requests and signed receipts. Each part of Round 3 meets a part of MeshKit, as FIG. 11 shows. The kit opens a MeshKit endpoint in front of your shop, so you do not rewrite the shop. MeshKit grows with the course; send bugs and improvements you find through the kit as PRs.
From week 12 your shop gets a second door, as FIG. 12 shows. Market buyers come in over A2A and ACP as they do now, and phone apps and other agents come in through the MeshKit endpoint.
| Week | Market and MeshKit |
|---|---|
| 9 to 11 (Round 3) | Pro-activeness and social ability while Round 3 runs. Round 3 results in week 11 |
| 12 | The kit opens a MeshKit endpoint in front of your shop. Tests attack it with tampered, replayed, over-budget and impostor requests |
| 13 | Agent addresses and introductions. Buyers find each other and propose first (group buys, making plans) |
| 14 | Cross testing. Call each other's shops from a phone app |
| 15 | Demo day |
The midterm from week-08 HW continues in the Round 3 market. The submission folder is submissions/<student-id>/week-08.
Round 3 midpoint and pro-activeness. Find the agents that moved first in the Round 3 data, and when speaking first pays.