Skip to content

RFC: multi-agent failure attribution and credit assignment #4220

@bug-ops

Description

@bug-ops

Summary

Umbrella RFC consolidating 4 research issues targeting failure attribution and credit assignment across multi-agent execution.

Consolidated Sources

Adjacent: #4152 (RL from orchestration traces, kept independent) — same data substrate but different angle.

Existing Coverage

  • /specs/010-security/010-4-audit.md — Cross-tool injection correlation, audit trail
  • /specs/009-orchestration/spec.md — VeriMAP predicate gate, AdaptOrch topology advisor
  • /specs/035-profiling/spec.md — Per-span tracing instrumentation

Question to Resolve

Define the failure attribution model for Zeph multi-agent execution: which signals (provenance graph, prefix monitors, contrastive credit, LIFE classification) should be captured, where they live (audit log vs telemetry vs subagent transcript), and how they drive subsequent runs (reward shaping, capability adjustment, planner feedback).

Acceptance

Failure-attribution data model proposed; mapping to existing audit/telemetry/subagent transcript surfaces; RL feedback hook identified (or deferred).

Metadata

Metadata

Assignees

No one assigned

    Labels

    P3Research — medium-high complexityepicMilestone-level tracking issueorchestrationTask orchestration / DAG schedulingresearchResearch-driven improvementsub-agentsSub-agent orchestration

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions