Agents cooperate better when they can see each other.

First, the intuition.

Two workers editing the same codebase can overwrite one another or repeat work. Shared workspace state becomes useful only when agents interpret relevant activity.

What the researchers actually did

PASC gives two coding agents structured visibility into peer commits, lifecycle events, and recent tool/test output in a shared Git workspace, and compares against silent two-agent and single-agent configurations.

What the results suggest

• The peer-aware pair outperformed a solo baseline in the tested Python subset of SWE-bench Pro.
• A silent two-agent baseline did not significantly improve over one agent.
• The authors report fewer destructive concurrent edits and better cost per resolved task than silent two-agent work.

How convincing is the evidence?

The silent-team ablation helps isolate the importance of observable coordination instead of merely increasing agent count. A replication package is publicly linked in the paper.

What remains uncertain?

The benchmark subset and tested models limit external validity. Resource budgets are not established as fully matched across all comparison axes.

Connections worth exploring

• Git commits can serve as a shared action space.
• Explicit peer awareness reduces coordination ambiguity.
• More agents can increase conflicts unless the shared state and ownership rules scale with them.