Skip to content

Latest commit

 

History

History
executable file
·
70 lines (49 loc) · 4.19 KB

File metadata and controls

executable file
·
70 lines (49 loc) · 4.19 KB

Pattern 08: Learning Measurement Agent

An agent that continuously assembles the evidence of whether learning is changing behavior — so measurement stops being an annual survey and starts being a feedback loop.

Use Case

Program-level measurement for L&D and people-science teams. Most learning measurement is completion counts plus a smile-sheet, produced quarterly, read by no one in time to act. This pattern reframes measurement as an always-on agent that gathers, triangulates, and narrates adoption and behavior-change signals.

Learner Need

This is the one pattern whose primary user is the program team, but the learner-facing contract still matters: "If you're going to measure my learning, measure something real, ask me rarely, and show me what you found."

Agent Role

An evidence assembler and honest narrator. The agent pulls from defined signal sources (xAPI streams, adoption telemetry, lightweight pulse questions, transfer check-ins from patterns 06–07), triangulates them, and produces plain-language readouts with confidence levels. It is not an individual-performance scorer — it reports on programs and cohorts, never on named people.

Interaction Loop

  1. Trigger — a reporting rhythm (biweekly readout) plus threshold events ("adoption of X stalled for 3 weeks").
  2. Observe — agent ingests the signal map: behavioral events (what people did), self-report (what people say), and outcome proxies (what the work shows).
  3. Act — agent produces a readout: what's moving, what's stalled, where signals disagree, and what it cannot know from the available data. The "cannot know" section is mandatory.
  4. Learner response — program team acts: redesign, targeted support, or a deliberate experiment.
  5. Adapt — the agent tracks whether the intervention moved the signal, building an institutional memory of what actually works in this organization.

Data Required

Data Source Sensitivity
Behavioral events (tool usage, xAPI statements, workflow telemetry) LRS, product analytics High — aggregate and de-identify before the agent sees it
Pulse self-report (2 questions max, monthly max) Learners Medium
Outcome proxies (cycle time, quality metrics, review outcomes) Business systems High — use with explicit stakeholder agreement
Signal map defining what counts as evidence Program team, up front Low

Risks

  • Surveillance by another name — behavioral telemetry aggregated carelessly is monitoring, whatever you call it. De-identify at the pipeline level, report at cohort level, and tell learners exactly what is collected.
  • Metric theater — the agent confidently narrates noise. Force confidence labels and the "cannot know" section; a readout without stated uncertainty should not ship.
  • Goodhart drift — once a signal becomes the target, it stops measuring learning. Rotate and triangulate signals; treat any single metric's sudden improvement with suspicion.

Evaluation Signals

Meta-signals — evidence the measurement loop itself works:

  • Program decisions cite the readouts (redesigns, kills, doublings-down traceable to evidence).
  • Time from "signal stalls" to "team responds" shrinks.
  • Signal sources disagree sometimes. If they never disagree, you're measuring one thing three ways.
  • Learners can accurately describe what's being measured about them.

Example Prompt

You are a learning measurement agent for [program]. You report on cohorts
and programs, never on named individuals.

You receive: behavioral events (de-identified), pulse self-report, and
outcome proxies, per the attached signal map.

Every readout must contain:
1. WHAT'S MOVING — signals showing change, with the evidence and a
   confidence label (strong / suggestive / weak).
2. WHAT'S STALLED — and for how long.
3. WHERE SIGNALS DISAGREE — self-report vs. behavior vs. outcomes, and
   your best hypothesis for the gap.
4. WHAT THIS DATA CANNOT TELL YOU — mandatory. Name the questions the
   current signals cannot answer.
5. ONE suggested experiment or intervention, framed as testable.

Plain language. No dashboard-speak. If the honest answer is "we cannot
tell yet," say exactly that.