The engineer and their agents

A task specifies the asset, useful service, objective, constraints, data validity and acceptable output. For A-01, define available load reduction during a stated interval while maintaining the approved temperature envelope. A request to “reduce consumption” leaves substantial ambiguity.

The agent’s output includes inputs, calculations, a proposed action, predicted consequences, stop conditions and proposal expiry. Missing information can produce a precise measurement request. The human decides changes to objectives and limits, recording them in a new task version.

Four functions are sufficient for a teaching team: acquire and validate data, calculate, challenge errors and prepare execution instructions. Different agents or separate stages of one system can perform them. Each function has inputs, outputs and acceptance criteria.

Verification independence depends on access to primary data and an independent calculation method. Two copies of a model repeating an answer provide limited protection against shared errors. The checker inspects original units, reproduces the central calculation and tests an adverse condition. A named owner retains responsibility for the asset.

Every measurement needs a timestamp, time zone, unit, source, quality flag and asset identifier. Distinguish measurement time from message arrival. A duplicate message retains the same identifier and represents the same physical event.

At 11:20 in A-01, the latest reading is from 11:00 and maximum data age is 300 seconds. The reading is 1,200 seconds old, so a mode-change proposal waits for refreshed data. A power value of 2,400 W normalizes to 2.4 kW. Converting a kWh value to average power requires the interval duration.

A read tool returns data. A calculation tool accepts a validated schema and returns a numerical result. An execution tool checks authority, permitted range, freshness and command identity. Availability of a function does not establish permission to control an asset.

Store command_id, asset, requested state, expiry, justification and recovery condition. Repeating a command identifier should return its first execution result. Text from an external document is treated as evidence. Instructions found within it cannot expand the agent’s assigned authority.

A 12 kW reduction for 15 minutes corresponds to 3 kWh shifted or avoided during that interval. If recovery adds 4 kW for one hour, the complete window shows a 1 kWh increase in consumption. The power service may still have value; an energy-saving claim requires the complete window.

The agent checks temperature, equipment limits and recovery. A decision links initial state, prediction, action and measured consequence. Record prediction errors together with their causes and corrections. Apply a teaching model within its explicit assumptions.

The evaluation set includes normal cases, boundaries and deliberately corrupted inputs. Define expected checks, acceptable numerical ranges and refusal reasons before running each task. Score numerical accuracy, evidence, appropriate stopping and adherence to authority.

Separate training examples from cases withheld until assessment. After a model or instruction update, rerun a stable regression set alongside new cases. Report the overall score together with critical failures. In this Academy exercise, a critical failure blocks progression to the corresponding action class.

Deployment progresses through historical replay, live-data observation with logged proposals, and limited application after technical acceptance. During observation, compare agent proposals with actual decisions and outcomes.

Each stage specifies duration, required operating conditions, progression criteria and recovery criteria. One elapsed week says little about coverage of a rare winter condition. Logs preserve inputs, model and tool versions, decision, time and outcome, allowing an event to be reconstructed after the system changes.

A useful experience becomes a verifiable case: conditions, decision, measured outcome and transfer limits. One successful week can produce a hypothesis for the next trial. An error can become a new evaluation example.

Collecting documents, updating a knowledge base, revising instructions and training model parameters are distinct operations. Each needs versioning and evaluation. The human learns to identify flawed assumptions and ask precise questions. The agent develops reproducible task performance. Joint progress is measured through decisions on new assets and the consequences of their use.

Prepare an agent task for one asset, input schema, tool and authority list, at least 20 evaluation cases, a logging protocol and a transition plan from historical replay to observation. Cover time, units, sources, duplicate execution and load recovery.

Give another evaluator ten new cases. Record numerical accuracy, justified stops and critical failures. The practical defence includes observing the agent on a real task within approved boundaries and analysing one case in which the original decision required correction.