Agentic Quality Intelligence should earn its complexity. This experiment asks whether coordinated agents actually improve assurance decisions—or simply create a more elaborate workflow.

The hypothesis

A coordinated set of specialized quality agents can make a better assurance decision than either one general-purpose agent or a fixed workflow—provided they share evidence, expose disagreement and operate inside explicit decision boundaries.

The experiment is not about proving that multiple agents can complete more tasks. That is easy. The harder question is whether coordinated reasoning improves the quality of the decision.

The decision we will test

Use one bounded question:

Given this release change and the available evidence, where should assurance effort be increased before release?

That decision is narrow enough to evaluate and broad enough to require several types of reasoning.

Experiment setup

Create a controlled release scenario containing requirements, change summaries, dependency information, historical defects, test coverage, recent execution, telemetry and business criticality.

Run the same scenario through three approaches:

  • Baseline: a deterministic rules-based prioritization.
  • Single agent: one general-purpose reasoning agent with all evidence.
  • Agentic QI: several bounded agents using shared evidence and an orchestration layer.

The agent roles

Change Intelligence Agent

Identifies what changed, which components are affected and where dependencies expand the blast radius.

Risk Agent

Combines change, history, business consequence and dependency exposure into explicit risk hypotheses.

Coverage Agent

Examines whether current test and evaluation evidence actually covers the identified risks.

Runtime Agent

Looks for telemetry, incident and performance signals that support or contradict the emerging risk position.

Confidence Agent

Builds the final recommendation, preserving the evidence and disagreement produced by the other agents.

Shared evidence is a requirement

The experiment fails architecturally if each agent builds its own private version of truth.

All agents should operate against a shared evidence model containing source identifiers, timestamps, versions and relationships. The point is not simply to share text. The point is to share the same quality state.

What happens when agents disagree?

Disagreement is one of the most important test conditions.

The Risk Agent may see elevated exposure while the Coverage Agent sees strong validation. The Runtime Agent may introduce a signal that weakens the confidence position. The orchestration layer should not hide this conflict by averaging the answers.

It should surface the disagreement and ask what additional evidence would resolve it.

What we measure

Decision relevance

Did the recommended assurance actions focus on the risks that later proved material?

Evidence efficiency

Did the system identify the smallest amount of additional evidence needed to improve the decision?

Explainability

Can a reviewer trace the final recommendation to the agents, evidence and reasoning behind it?

Consistency

Does the system make similar decisions for materially similar release conditions?

Coordination overhead

Does multi-agent orchestration create enough latency, complexity, or token costs to outweigh the improvement in decision quality?

What would falsify the hypothesis?

The experiment should be considered unsuccessful if the multi-agent approach produces more elaborate explanations without materially better decisions, if agent disagreement becomes orchestration noise, or if a single well-grounded agent performs equally well with much lower complexity.

Expected failure modes

  • agents duplicate the same analysis under different names;
  • weak shared context causes contradictory conclusions;
  • the orchestrator becomes a hidden super-agent that does all the reasoning;
  • confidence scores create false precision;
  • agents optimize their local objective rather than the release decision;
  • human reviewers cannot reconstruct why the final recommendation changed.

The next experiment

If coordinated agents improve prioritization, increase the level of autonomy slightly. Allow the system to execute one reversible action—for example, expanding targeted regression or gathering additional telemetry—without human approval.

Then evaluate whether the action stayed inside policy and actually improved the confidence position.

The interesting question is not whether agents can test software. It is whether several bounded forms of intelligence can make a better quality decision together.