Topic 8 of 20 · Part II : Grounding, Evaluation and Orchestration

Agent systems

An agent's behaviour depends on the data describing its environment : tools, observations, permissions and workflow state become knobs for autonomy and risk tolerance.

The control pattern

Data asset

Tool, state and workflow data

Knobs

Available tools, Permissions, Memory, Planning depth, Stop conditions, Action budget

Outcome

Better multi-step decisions with controlled execution

Measurement

See how to measure this below

What it is

An agent is an AI system that can select tools, maintain state and take multiple steps toward a goal. Its behaviour depends heavily on the data describing its environment: available tools, prior observations, user intent, permissions and the current workflow state.

These become knobs for how autonomous, persistent and risk-tolerant the system should be. A useful agent has explicit stop conditions and budgets. It should not continue gathering data or taking actions indefinitely simply because more steps are possible.

Distillation can later compress stable agent behaviours such as query planning or result summarisation, but the permission and audit layers should remain explicit rather than being hidden inside model weights.

The knobs in detail

Each row is one adjustable property of the data asset, and what moving it tends to do.

KnobWhat you adjustLikely effect
Available toolsWhat the agent may callBounds what it can possibly do
PermissionsRead-only vs. consequential actionsSeparates research from execution
MemoryWhat state persists between stepsControls coherence and leakage
Planning depthHow far ahead it reasonsTrades thoroughness against cost
Stop conditionsWhen the agent must haltPrevents indefinite loops
Action budgetMaximum calls or spend per taskCaps blast radius and cost

Applied: Stocks Assistant

A Stocks Assistant research agent could resolve a ticker, retrieve the latest filing and transcript, extract key metrics, compare them with prior periods and peer ranges, check market data, and assemble a cited briefing. A portfolio-review agent could do the same for each holding and aggregate concentration or catalyst risk. Each tool result should be stored with timestamp and provenance. Read-only research can be broadly automated, while consequential actions : placing a trade, modifying a paid subscription, sending a personalised alert : require explicit permissions and confirmation.

How to measure it

Evidence that the knob produced the intended behaviour, rather than shifting the problem elsewhere.

  • Task completion and evidence coverage
  • Unnecessary tool calls
  • Failure recovery behaviour
  • Policy compliance
  • Replayable traces showing why each tool was chosen

Common mistakes

Automating consequential actions on the same footing as research.

Agents without replayable traces, which cannot be debugged or audited after the fact.