Provenance Layer

I build systems that show where AI outputs and agent actions came from.

My recent work sits around one system boundary: source context becomes generated knowledge, tool calls, product decisions, and audit logs. That path needs provenance.

Agent Runtime InspectorAI integration layer article

What I Mean

Source context

The system records which source messages, documents, tickets, or transcripts fed the AI workflow.

Permission boundary

The system carries access rules forward when raw source data becomes a summary, decision, answer, or tool argument.

Runtime path

The system records which tools were visible, which tool was called, what arguments were sent, and what result came back.

Audit record

The system leaves a readable trail that an engineer, operator, or security reviewer can inspect after the run.

Work

Sift decision traces

Private platform work

At Sift, I worked on systems that traced derived decisions back to the source context that produced them. The trace model carried source IDs, source types, timestamps, authors, channels, and roles such as Proposal, Context, Objection, Validation, Pivot, and Resolution.

Replaced flat underlying-discussion data with chronological decision traces.

Connected decisions to source messages, conversations, and meeting transcripts through graph relationships.

Scoped decision access by organization, user, and private-channel membership.

Treated provenance as part of permissioning, because private source content can leak through derived AI knowledge.

View reference

Agent Runtime Inspector

Public OSS project

ARI applies the same provenance idea to AI coding agents. It sits between an MCP-capable agent client and Merge Agent Handler, forwards tool calls, and records what happened in a local dashboard.

Records the tool inventory exposed to the agent.

Records the selected tool, arguments, result, latency, and errors.

Separates the run into context path, action path, and model path.

Supports mock traces, scripted Merge examples, and an MCP proxy path for local agent clients.

View reference

Where This Fits

AI coding agents that call GitHub, Linear, Slack, Jira, or internal tools
Enterprise agents that need per-user tool access and audit trails
Knowledge assistants where answers must preserve source permissions
Developer tools that need to explain what an agent changed and why
AI infrastructure teams building observability, governance, or integration layers

If your team is building agents that touch customer systems, I can help with provenance, permission-aware retrieval, tool execution, integration architecture, and developer-facing explanation.

Start a conversation