Where your agent runs

An agent is only as governed as the runtime it runs on. These are the six targets Scalarion places into, the rules that choose between them, and — the part nobody publishes about their own product — exactly which guarantees each runtime enforces, which Scalarion enforces on top, and which nothing enforces at all and a person still owns.

The placement below is worked from a sample assignment so the rules are visible without an account. Describe your own agent and the same rules run against it.

SAMPLE — a worked example, not your workflow. Run the audit and this fills with your own map.

TARGET PLACEMENT · SCALARION RECOMMENDS · YOU CAN OVERRIDEfirst rule that fires, wins

RULE 5 FIRED

2 agent step(s) across 6 mapped, 2 observed at volume — judgment that must stay auditable step by step.

THESE RULES NEED A FACT THE AUDIT DOESN'T HOLD — ANSWER ONE AND THE PLACEMENT CHANGES

Rule 2 → Agentforce: Does this work live entirely inside one system of record with a native agent platform?
Rule 3 → Copilot Studio: Is the workspace itself the workflow — email, chat, files, meetings?
Rule 4 → AWS AgentCore: Is execution long-running, isolated, or high-blast-radius?
Rule 7 → LangGraph: Are you sovereignty-constrained or air-gapped? (self-hosted, local models)

GOVERNANCE PARITY — LANGGRAPHFULL PARITY

THE RUNTIME ENFORCES

— typed state — every transition declared
— checkpointers — replayable history
— native interrupts — pause at any node

SCALARION ENFORCES ON TOP

— PEP policy check at every tool node
— interrupt = HOLD on each danger step
— checkpoint stream → sealed ledger events
— eval harness wired to the golden set

NOBODY ENFORCES → STAYS HUMAN

— money and irreversible actions stay held — or the agent doesn't ship: one that can move money unattended is flagged, forced to no-go, and refused at the deploy gate
— approving any policy exception

The third column is the build phase's honest no — no vendor will ever publish it about themselves. It prints verbatim on the emitted artifact.

CEILING ON THIS TARGET

SHADOW

Unsealed — the artifact prints est. on its face. Seal the run in 4.1 to raise it.

UNHAPPY PATH — DESIGNED BEFORE THE HAPPY ONE

E5 Runtime outage mid-run — the target goes away after partial execution

governed response: checkpoint replay · the run is marked PARTIAL on the ledger — a partial run is never sealed.

Emission is free — the manifest is the asset, the seal is the paid proof. Re-emit to a second target any time: same manifest, evals re-run, divergence logged as a finding.