Skip to content

Execution tuning (not an Agent field)

This page tells you where engine tuning goes. spec.execution on kind: Agent and spec.patch.execution in an overlay are not supported. Put engine queue depth, cache policy, streaming, and parallel tool dispatch in a kind: RuntimeEngine manifest.

The runtime resolves a RuntimeEngine the same way as infra. It looks in workspace runtime_refs, then in $XDG_CONFIG_HOME/mas/runtime/, or takes the CLI flag --runtime-ref.

Smallest replacement for an old spec.execution block:

apiVersion: infra/v1
kind: RuntimeEngine
metadata:
  name: my-runtime
spec:
  engine:
    queue_depth: 64
  stream: false

See runtime-engine.md and user-config.md.

Migration table

Old (invalid on kind: Agent / overlay patch) Replacement
spec.execution (cache, stream, queue depth, …) kind: RuntimeEngine via runtime_refs / --runtime-ref
Offline LLM turns llm_cache replay recorded against a live provider (raise_on_miss)
spec.infra_refs / spec.runtime_refs on Agent, MAS, or overlay config.yaml, --infra-ref / --runtime-ref
spec.infra_interceptors on Agent workspace / CLI infra resolution (same as infra refs)

Do not confuse with experiment settings

Batch orchestration for mas-lab benchmark run lives on the experiment manifest, not on agents. The deprecated experiment.execution block is split into three keys:

  • experiment.design: what the experiment compares.
  • experiment.schedule: parallelism, timeouts, ordering, and retries.
  • experiment.bench_emulation: the trace-cache posture (how runs reuse stored traces).

None of them replaces RuntimeEngine. See Design vs schedule.

mas-ctl validate and the separation checks reject the old keys, so outdated examples fail at once.

See also