Runtime makes the agents run. Scenario architecture makes the run interesting.
Do not design a pipeline. Design pressure.
roles + private information + goals + actions + consequences + artifact history
=> emergence
A strong WorldSeed scene has these parts:
roles who acts
private info what each role knows or cares about
goals what each role tries to protect or win
actions what agents can do to the world
consequences who gets woken, blocked, pressured, or rewarded
artifact history what remains visible after the run
If the scene only says "producer writes, critic critiques, curator selects", it will behave like a normal multi-agent pipeline.
Roles should create pressure, not just divide labor.
Examples:
builder wants its branch to survive
critic wants to reject boring or unsupported work
audience protects attention and clarity
technical protects reproducibility
curator must ship and justify the final package
The point is not conflict for drama. The point is better artifacts through different incentives.
Do not give every agent the same view.
Examples:
API explorer sees constraints and failure modes
prompt miner sees patterns and anti-patterns
audience sees usefulness and confusion
technical judge sees reproducibility risks
curator sees the whole artifact graph
Asymmetry is what makes agents disagree for real reasons.
An action should not only record text. It should change what happens next.
Examples:
submit_version creates a candidate and wakes judges
critique_artifact attaches pressure to one exact artifact and wakes owner
select_branch changes artifact status and wakes owner
submit_package closes a run with cited evidence
If actions do not create consequences, the scene becomes a shared folder.
Emergence should be inspectable after the run.
Use append-only records for:
attempts
versions
critiques
rebuttals
revision requests
selections
final packages
Final output should cite exact artifact ids, including rejected or revised branches. The trail is part of the product.
The primary session should not pre-write the whole flow.
It should:
start with enough structure
spawn parallel branches when useful
read signals and artifact ids
wake agents when consequences require reaction
increase pressure when outputs are generic
stop when the artifact graph is strong enough
Good runs feel designed at the rules level, not scripted at the outcome level.