Coordinate agents with limited information.

This is the same reference used in Technical mode. Expand it whenever you need the exact numbers, assumptions or further explanation.

Each teammate has only half the address

Two agents must meet at a corner of a 3×3 grid. Agent A sees the target row, while agent B sees its column. A starts at row 1, column 0; B starts at row 1, column 2. Knowing that there are two agents does not magically combine their observations. The information has to reach the decision that needs it.

The shared success condition requires both agents at the true target within six ticks. Arriving separately at different guesses fails, even if each individual followed its own rule correctly. This distinction between component behavior and complete task success applies to software workflows as well as simulated robots.

Three observation contracts

In local mode, each agent fills its missing coordinate with the middle coordinate, 1. A target at row 0, column 2 therefore looks like (0, 1) to A and (1, 2) to B. Neither local target is the true corner. The error is missing information rather than a faulty movement implementation.

Centralized mode supplies both coordinates to both controllers. Communication mode explicitly sends the missing coordinate in each direction before movement: two messages. It obtains the same task information as centralized control but incurs a different communication resource count. The experiment does not invent network latency or convert two messages into a guaranteed wall-clock time.

Each controller moves toward its believed target, changing row first and column second. This is a programmed policy. No neural network, policy gradient or learned messaging protocol is hidden in the demonstration. Multi-agent learning is an optional extension, not a claim established by this observation comparison.

Why action timing matters

At every tick, both actions are computed from the same pre-tick states. Then both next states commit together. Agent B does not get to see A’s newly committed position before choosing its own action. The trace explicitly shows both before/after states so this timing is inspectable.

Sharing a cell is allowed. The agents do not block one another and there is no collision penalty. These are environmental assumptions, not facts about physical robots. Changing them would change the task and require a new protocol, rather than merely another policy label.

Inspect contribution without inventing causality

Disable A’s movement and compare the shared success result. This reveals that reaching the joint target requires both agents in this task. It does not isolate a learned causal credit-assignment signal: the intervention changes team capability and the policy never learns from the result.

Inspect all four corner targets instead of selecting only a successful example. Compare completion, ticks and messages separately. A controller can spend fewer messages by failing to obtain necessary information; that is not a useful efficiency gain under the shared completion contract.

The four corner scenarios are public development cases. They do not establish robustness to delayed messages, adversarial teammates, larger maps or real-world movement. A proposed extension should name its observation, action, timing and success contracts before collecting outcomes, then retain failures alongside successes in Project 14.