Add stage-aware action receipts and uncertainty capture for collaborative agent execution

[tcconnally]

Source

What 5M+ daily MCP tool calls taught us about the future of AI at work

Why this matters

The article focuses on large-scale real-world agent usage and highlights the operational patterns that make AI systems dependable in practice. The most relevant themes are previewing actions, decomposing work into inspectable stages, grounding actions in live systems of record, and designing around human trust.

Key takeaways to apply

• action plans should be inspectable before execution
• complex work should be decomposed into auditable stages
• systems should preserve grounding to live source-of-truth data
• human trust improves when uncertainty and action boundaries are explicit

Proposed repo target

Ledger

Problem statement

Ledger has strong action-intent and receipt primitives, but it can push further toward stage-aware collaborative execution. We should make action planning, approval, execution, and uncertainty state more explicit and auditable as first-class runtime records.

Proposal

Extend Ledger with stage-aware execution receipts and uncertainty capture for agent actions.

Potential scope:

• model multi-stage action flows: proposed, approved, leased, executing, completed, failed, canceled
• attach source-context hashes and policy hashes to each stage
• record machine-estimated uncertainty or confidence state where available
• support human intercept points with deterministic approval receipts
• emit compact artifacts that upstream context systems can display before actions run

Success criteria

• every write-capable action can produce a complete stage trace
• action receipts can reconstruct what was proposed, approved, and executed
• upstream agents can surface a concise execution preview before mutation steps

Suggested labels

enhancement, reliability, auditability, mcp

Citation mapping

This issue is directly motivated by the operational lessons described in the large-scale MCP usage article.