What did my AI agent actually do?

The question behind the question

When someone asks what an agent did, they usually want six things: the instruction it received, the steps it took in order, the models and tools involved, what each step produced, where the time went, and what failed or was retried.

Chat transcripts answer none of these reliably. They show the conversation, not the execution. A single reply can hide a dozen tool calls, three retries, a web fetch and a file write.

What an execution record contains

• The instruction that started the run, as the runtime received it.
• Every step in execution order, with its own start and end.
• The models, tools, services and sub-agents involved in each step.
• Outcomes: completed, failed, interrupted, retried or recovered.
• Duration for the run, and separately where the activity time went.
• The points where a human intervened, approved or overrode.

What is missing matters as much as what is there

An honest record says when something was not captured. Inficy shows uncaptured facts as unavailable rather than as zero, because a zero implies a measurement that never happened.

Capture coverage is itself a metric. If a runtime cannot report token counts or step durations, that gap belongs on the page, not hidden behind an average.

How Inficy records it

Supported runtimes, including Codex CLI, Claude Code and Hermes, are observed at runtime and preserved as canonical NexArt Project Bundles. Inficy verifies those bundles for integrity, projects them into readable timelines and analytics, and lets you export them.

ChatGPT connector sessions are different: they are checkpoints the assistant reports itself over MCP. They are useful for tracking activity, but Inficy did not observe the underlying execution, so they are labelled as agent-reported throughout.