Palantir Begins Tracking Agent Work Hours—Measuring Enterprise AI in the Observability Era

On August 11, Palantir added a capability called Object Timeline to its Foundry/AIP stack. Although the name suggests a simple change‑log, the timeline links agents, human actions, executions, token consumption, runtime, waiting time, and a new metric called Agentic Coverage into a unified work‑time line for each business object.

This shift answers the most pressing post‑deployment question for enterprise agents: not merely whether a model can finish a task, but how much of the end‑to‑end workflow is actually driven by the agent, where delays occur, what resources are consumed, and which steps still rely on humans.

The top of the timeline displays four summary metrics. Total token usage shows the cumulative tokens spent by all executions on the object; Total runtime measures the elapsed time from the first to the last execution; Waiting time records how long the process paused for human input; and Agentic Coverage indicates the proportion of the object’s history driven by agents rather than people. In Palantir’s official example, a marketing campaign consumed 5,427 tokens, ran for 57 minutes 39 seconds, spent 98 % of that time waiting for human input, and achieved an 86 % Agentic Coverage (6 / 7 automated).

While Agentic Coverage is useful, the article warns against interpreting it as a direct labor‑saving percentage. A high coverage figure does not guarantee correct outcomes, and a low figure may be intentional for high‑risk approvals where human oversight is required. The metric should be considered alongside quality, cost, and the reasons for any human intervention.

Traditional LLM/Agent observability focuses on trace‑centric data such as calls, spans, tool usage, token counts, latency, and error rates. Object Timeline moves the focus to the object‑centric level, showing the full lifecycle of a business entity (e.g., a campaign, order, or device ticket) and how agents and humans jointly modify its state.

Consequently, the evaluation framework for enterprise agents expands from pre‑deployment benchmarks (success rate, accuracy, Pass@1) to post‑deployment layers: efficiency (token and cost), automation (Agentic Coverage), collaboration (waiting time and hand‑off points), and responsibility (traceability of each change to a specific execution).

These metrics also reshape optimization priorities. If the timeline reveals that 98 % of time is spent waiting for human input, improving model speed yields little benefit; instead, teams should streamline hand‑offs. If a single agent dominates token consumption, its context size, routing logic, or task decomposition should be examined. High coverage paired with frequent rollbacks signals quality issues rather than insufficient automation.

Object Timeline is currently part of Palantir Autopilot and remains in beta, but it illustrates a broader industry move: from monitoring model internals to monitoring the actual work performed by agents and humans together. The next generation of agent observability will likely treat “work” as the primary unit of measurement, not just model calls.