What you’ll make work
Change an energy-only evaluator to check class coverage too. Compare schedules, repair missing capacity, and verify a feasible improvement.
New to AI agents? Start with seven short foundation lessons on models, tools, feedback, evidence, and boundaries. Then return here to apply the ideas.
Each run connects a goal to a tool result, a check, and a next action. You change a meaningful setting, watch what changes, then apply the idea to a new case.
- Goal
- Reduce simulated energy use across six one-hour slots.
- Tools
- Simulate the schedule, validate capacity, and compare feasible candidates.
- Check
- Supply all 20 devices during required class slots 2, 3, and 5.
- Stop
- Return a feasible schedule and its checked energy total.
This is a guided simulation: its authored controller selects from a bounded set of actions, while its tools execute calculations and checks. A live model would choose actions dynamically.
Original worked example · fictional scenario
A zero-energy result that fails the job
This fictional model has 20 identical devices and six one-hour slots. Each powered device uses exactly one simulation energy unit per slot. Startup and standby costs are zero by definition. All 20 devices are required in slots 2, 3, and 5.
| Candidate | Devices in slots 1–6 | Energy units | Class coverage |
|---|---|---|---|
| Baseline | 20, 20, 20, 20, 20, 20 | 120 | Pass |
| All off | 0, 0, 0, 0, 0, 0 | 0 | Fail · 3 classes missed |
| Repaired | 0, 20, 20, 0, 20, 0 | 60 | Pass |
simulate_schedule computes the total powered device-slots. validate_capacity rejects the zero-energy schedule. compare_feasible chooses the repaired 60-unit schedule over the 120-unit baseline.
What happens nextThe agent keeps the 60-unit schedule because it serves every class. If slot 4 additionally requires 10 devices, it revises that slot and returns 70 units with all requirements met.
A score is useful only when it represents the actual job. Minimizing energy alone rewards switching everything off, even when that prevents classes. First reject schedules with missing capacity; then compare their energy use. Under this model, three required slots × 20 devices × 1 unit establishes a 60-unit lower bound. That is a calculation from the assumptions, not the agent’s confidence about its answer.
Common mistakes worth catching
Calling the lowest score a success
The all-off schedule has the smallest energy value and fails every required class. Check feasibility before comparing scores.
Treating partial capacity as coverage
Ten devices in slot 3 leave a shortage of 10. The fixture requires 20, so partial coverage fails that requirement.
Treating simulation units as measured electricity
Actual equipment has additional behaviors and costs. These explicit assumptions define a teaching model, not a physical energy forecast.
What to take with you
- An evaluator must measure the goal and required constraints.
- Reject infeasible candidates before ranking by efficiency.
- Use the model’s assumptions to calculate and explain a valid bound.
Finishing the activity gives you a record of the configuration you changed, the case you tested, and the result you observed. The new case checks your understanding separately from the guided run.
A few good questions
Why is 60 the minimum here?
Three required slots each need 20 devices, and each powered device costs one unit per slot. Those required slots alone total 60 units; the repaired schedule adds zero elsewhere.
Is this how much power a real lab saves?
No. The unit is defined for the simulation. Real energy use depends on hardware, workloads, standby behavior, startup costs, and other conditions absent from this model.
What changes when a new class is added?
The validator gains a new capacity requirement. If slot 4 needs 10 devices, the lowest feasible schedule in this fixture consumes 70 units rather than 60.
Research behind this lesson
Google DeepMind · May 14, 2025
Generates candidate programs and uses executable evaluators to score outcomes and guide improvement. AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms ↗
Our six-slot schedule borrows candidate → evaluator → revision. It is not AlphaEvolve and does not model real equipment energy behavior.
This school scenario is a simplified simulation. The research connection explains the method that inspired the activity; it does not establish this lesson’s effectiveness or imply an endorsement.
Content version: 2026-09-12.1 · Prepared September 12, 2026
Editorial review pending · No completed specialist review or review date recorded.