Detailed Analysis
Lane Harlan, a founder and former SAFe Agile practitioner with a background in pharmaceutical IT program management, has proposed a new unit of measurement for agentic AI work called PACE — a backronym for Per Agent Compute Estimate. The unit is calibrated to 100 tokens per second sustained over one hour, equating to 360,000 tokens per PACE. Unlike raw token counts or human-duration estimates, PACE introduces a time dimension designed to make capacity planning and sprint forecasting tractable for teams deploying LLM agents as workers rather than human contributors. Harlan frames the proposal explicitly as "Agile for Agents" — an attempt to give AI-driven workflows the same kind of planning vocabulary that story points gave human software teams in the early 2000s.
The motivation behind PACE stems from practical failures Harlan encountered with existing estimation approaches. Asking an agent to self-estimate yields human-duration answers ("about two weeks") because the model's training data reflects human work, not its own computational throughput. Estimating in raw tokens is more technically honest but requires repetitive back-of-envelope conversion math that doesn't scale across projects. Story points, the dominant currency of Agile planning for human teams, are anchored to perceived cognitive complexity — a dimension that maps poorly onto agent compute. The only comparable commercial concept Harlan found was Salesforce's Agentic Work Unit (AWU), mentioned in earnings disclosures, which appears proprietary and operates at a granularity of roughly 8,000 tokens — too fine-grained for high-level planning purposes and unavailable as an open standard.
The practical utility of PACE as Harlan describes it lies in its model-agnostic translation layer. Using published throughput benchmarks, a PACE can be converted into approximate wall-clock duration for any given model: Claude Sonnet, which generates tokens at roughly 100 tokens per second, produces approximately one PACE per hour, while Claude Opus, operating at roughly half that speed, produces approximately one PACE every two hours. This means a 20-PACE task estimate yields an immediate duration forecast — 20 hours on Sonnet, 40 hours on Opus — without requiring the estimator to re-derive throughput constants each time. The analogy Harlan invokes is the kilowatt: just as a kilowatt meaningfully combines a rate (watts) with an implicit time context for energy budgeting, PACE combines token throughput rate with an hourly time anchor for compute budgeting.
Harlan is careful to separate the existence of the unit from the problem of estimation accuracy. He explicitly acknowledges that agents cannot yet reliably self-estimate work in any unit, and that PACE does not solve that problem. The argument is instead that a shared, consistent vocabulary must exist before estimation accuracy can improve — echoing the historical trajectory of story points, which were initially arbitrary but became meaningful through team calibration and shared convention. Crucially, Harlan argues that PACE would carry a fixed, cross-organizational meaning anchored to a physical throughput rate, unlike story points, whose meaning drifts between teams and companies. This positions PACE as potentially more interoperable as an industry standard than story points ever were.
The proposal surfaces a genuine and underexplored gap in the emerging field of agentic AI operations. As organizations move from treating LLMs as single-query tools to deploying them as persistent workers within automated pipelines, the absence of planning-compatible units of work becomes a real bottleneck for project managers, product owners, and engineering leads attempting to build roadmaps and allocate capacity. Harlan's open, copyright-free release of PACE reflects a pattern common in the early Agile movement — practitioners developing and sharing practical tools ahead of formal academic or vendor standardization. Whether PACE gains traction will depend on community adoption and whether model providers begin publishing throughput benchmarks in forms compatible with planning frameworks, but the underlying problem it addresses is likely to grow more acute as agentic deployment scales across enterprise environments.
Read original article →