Doc. TMF-005 · Rev 2026.08 Service Specification Classification: Open Sheet 5 of 6

Service Spec · Agent Evals & Assurance

Prove your agents work. Then prove it tomorrow.

Almost every organisation shipping agents can demo one. Far fewer can tell you whether today's release is better or worse than last week's. We build the harness that answers it.

Section 01 · Definition

Why agents need a different kind of testing

Conventional tests assume that identical input yields identical output. Agents break that assumption, so the usual green tick tells you almost nothing. Assurance replaces it with measurement: curated golden sets that encode what good looks like, graded rubrics calibrated against human judgement, adversarial suites that probe for injection and tool abuse, and production tracing that catches the drift no offline suite predicted. The result is a number that moves when quality moves, and a gate that stops a regression before your customers find it.

Fig. 01 Offline evals, adversarial suites, and production tracing in one loop.

The Equation

Golden sets
+
Adversarial suites
+
Regression gates
=
Shippable agents

Section 02 · Benefits

What you get

Ship on evidence, not vibes

A score that moves with quality, so release decisions stop being an argument about who tried the scarier prompt.

Swap models without fear

When a new model lands, run the suite. Change becomes a measurement instead of a gamble.

Find the attacks first

Prompt injection, tool abuse, data exfiltration, and excessive agency, probed deliberately before somebody else finds them.

Control cost and latency

Token spend and response time treated as budgets with thresholds, not a surprise on the monthly invoice.

Catch drift in production

Traces and online signals surface the degradation that no offline set anticipated, because reality supplies inputs you did not imagine.

Evidence for the auditor

When a regulator asks how you know the system behaves, the answer is a suite, a history, and a trail.

Section 03 · Scope

Our comprehensive offering

Evaluation harness

  • Golden sets and task suites built from your real traffic
  • Rubric grading, with model-as-judge calibrated against human labels
  • Per-capability scoring so you can see which behaviour regressed
  • Statistical treatment that distinguishes a real change from noise

Adversarial and security testing

  • Prompt injection, direct and via retrieved documents
  • Tool abuse, privilege escalation, and excessive agency
  • Data exfiltration and prompt or context leakage
  • Jailbreak suites maintained as new techniques emerge

Production assurance

  • Distributed tracing of every agent run, tool call, and retry
  • Online experiments and shadow deployments
  • Drift, cost, and latency monitoring against SLOs
  • Regression gates wired into CI so a bad build never ships

Section 04 · Engagement Model

Two ways to fly

Build and Transfer

We build the capability and then hand it over to your team, documented and running.

Build and Operate

We build it and keep running it, on a fixed-price subscription.

Pricing

Scalable without sacrificing affordability

Tiered on the number of agents, models, and environments under assurance, on the same fixed-price subscription model as our DevSecOps offering.

Fig. 02 Pricing tracks the estate under management.

Section 05 · Technology

The stack we reach for

Evals & experimentation
Golden setsLLM-as-judgeA/B and shadowHuman calibration
Test engineering
pytestPlaywrightHypothesisSchemathesisk6
Tracing & telemetry
OpenTelemetryPrometheusGrafanaTempo
Delivery
GitHub ActionsGitLab CIArgo CDMLflow

Section 06 · Why Us

Why choose The Moonshot Factory

Assurance is a testing and observability problem wearing new clothes, and both are already ours. We instrumented an entire microservices trading platform with OpenTelemetry and gave it a single pane of glass for metrics, logs, traces, and service map. We run DevSecOps as a subscription, with automated compliance checks inside the pipeline. Agent assurance is the same discipline pointed at a non-deterministic system.

Get Started Today →

Fig. 03 Same practices we preach, on every project we run.

Section 07 · Contact

Want to build a rocketship?

Contact Us →