In short: Agent-to-agent testing is a method where autonomous evaluator agents hold thousands of realistic conversations with your AI agent and score every response for accuracy, hallucination, bias, toxicity and context retention. It exists because traditional test automation cannot validate a system that answers differently every time it is asked the same question. Pro Skills delivers agent-to-agent testing across Saudi Arabia in partnership with TestMu AI, formerly LambdaTest.
What is agent-to-agent testing?
Agent-to-agent testing is a testing approach in which two or more AI agents interact with each other in a controlled environment to simulate real-world conditions. One agent acts as the evaluator, playing the role of a customer, a caller or an adversary. The other is the system under test. The evaluator scores how well the target agent understands, responds and holds context across a full conversation, not a single reply.
It answers a question that scripted automation cannot: does this agent behave correctly when nobody has written the script in advance?
The term also carries a second meaning worth knowing. Google's Agent2Agent (A2A) protocol, released as an open standard in 2025 with more than fifty partner organisations, defines how agents from different vendors discover each other and delegate tasks using Agent Cards and structured JSON-RPC messaging. Where A2A governs how agents talk to each other in production, agent-to-agent testing governs whether what they say can be trusted. Enterprises adopting one almost always need the other.
Why traditional test automation breaks on AI agents
Selenium, Cypress, Appium and Playwright were built on a simple assumption: the same input produces the same output. Click a button, expect a known result. That assumption holds for a checkout page. It collapses the moment the system under test is a large language model.
Five things break at once:
| Traditional application | AI agent |
|---|---|
| Deterministic output | Probabilistic output that varies per run |
| Fixed UI locators | No UI at all, or a conversational one |
| Pass or fail assertions | Quality exists on a spectrum |
| Finite test paths | Effectively infinite conversation branches |
| Failures are visible errors | Failures are confident, fluent and wrong |
That last row is the dangerous one. A broken form throws an error. A hallucinating banking assistant quotes a profit rate that does not exist, in perfect Arabic, with total confidence. Nothing in a conventional CI pipeline catches it.
Gartner expects that by 2028, at least 15% of day-to-day work decisions will be made autonomously through agentic AI, up from effectively zero in 2024. The volume of decisions moving to agents is rising far faster than the discipline of validating them.
What actually needs to be tested in an AI agent
A useful evaluation covers nine dimensions. Teams that test only the first two ship the rest as risk.
- Task accuracy. Did the agent complete what the user actually asked for?
- Hallucination. Did it invent policies, prices, dates or entitlements?
- Context retention. Does it still remember turn three by turn twelve?
- Bias. Do responses shift by gender, nationality, dialect or name?
- Toxicity and safety. How does it behave when provoked, insulted or baited?
- Tool and handoff correctness. Did it call the right system, pass the right parameters, escalate to a human at the right moment?
- Prompt injection resistance. Can a user talk it out of its own instructions?
- Latency under load. Does quality degrade when a thousand callers arrive at once?
- Language and dialect handling. Does it hold up in Arabic as well as English, and across regional dialects?
For voice agents, add interruption handling, background noise, accent variation and silence recovery. For multi-agent systems, add coordination failures, where each agent is individually correct and the workflow is still wrong.
How agent-to-agent testing works in practice
The evaluator agents do the work a QA team could never do at volume.
Persona simulation. Evaluators adopt profiles: the impatient customer, the confused first-time user, the non-native speaker, the enterprise procurement lead. Each drives the conversation differently.
Adversarial probing. A second class of evaluator actively tries to break the agent, pushing it towards unsafe advice, off-policy commitments, prompt injection and data disclosure.
Scored, not asserted. Every conversation returns a score across the quality dimensions rather than a binary pass or fail, so teams can set thresholds and track drift release over release.
Pipeline integration. Evaluations trigger from the terminal or CI, so agent quality gates sit alongside unit and integration tests rather than in a separate manual review at the end.
TestMu AI's platform runs this at enterprise scale: autonomous evaluators deployed against chatbots, inbound and outbound voice callers and image analysers, with more than fifteen specialised testing agents, over two hundred voice profiles, and results scored across nine quality dimensions and pushed into existing CI/CD workflows.
Why Pro Skills partnered with TestMu AI
TestMu AI rebranded from LambdaTest on 12 January 2026, marking its shift from a cloud testing grid to a full-stack agentic quality engineering platform. Existing accounts, scripts, API keys and integrations carried over unchanged.
The platform is used by more than three million developers and testers and over eighteen thousand enterprises including Microsoft, OpenAI and Nvidia. It has been named a Challenger in the 2025 Gartner Magic Quadrant for AI-Augmented Software Testing Tools and recognised in the Forrester Wave for autonomous testing platforms. Alongside agent evaluation, it runs conventional automation across a real device cloud of more than ten thousand devices and three thousand browser combinations, and its KaneAI agent authors end-to-end tests from plain-English instructions, exporting to Selenium, Playwright, Cypress and Appium.
That combination matters for Saudi enterprises. Most organisations in the Kingdom are not testing agents instead of applications. They are testing a mobile app, a customer portal, a set of APIs and a new Arabic voice agent, and they want one platform and one partner across all of it.
Delivering testing automation in the Kingdom
Pro Skills delivers testing automation across Saudi Arabia in partnership with TestMu AI, combining platform capability with local implementation, enablement and support.
Our engagement covers:
- Quality engineering assessment. Current test coverage, automation maturity, release bottlenecks and agent inventory.
- Platform deployment and integration. TestMu AI configured against your environments and wired into Jenkins, GitHub Actions, GitLab CI, Azure DevOps or your existing pipeline.
- Agent evaluation design. Persona sets, adversarial suites and scoring thresholds built around your actual use cases, in Arabic and English.
- Automation migration. Existing Selenium, Appium and Cypress suites moved onto the grid, with flaky tests identified and stabilised.
- Team enablement. Structured training so your QA engineers own the system, delivered through the same organisation that runs the Kingdom's BICSI, PECB and AI Certs programmes.
- Managed operation. Ongoing suite maintenance, regression cycles and release-gate reporting where teams prefer to outsource the run.
This sits inside our wider Agentic Software Testing platform practice, alongside agentic software development and AI consulting.
Why this matters for Vision 2030
Saudi Arabia's digital transformation is putting AI agents into banking, telecom, healthcare, government services and giga-project operations. Those deployments carry obligations: SDAIA's AI Ethics Principles set expectations around fairness, reliability, transparency and accountability, and the Personal Data Protection Law governs how customer data is handled in the systems these agents touch.
Bias testing, hallucination detection and audit-ready evaluation records are no longer engineering niceties. They are the evidence that an agent was fit to deploy. Agent-to-agent testing produces that evidence as a by-product of the release process rather than as a separate compliance exercise.
A practical first 30 days
For teams with an agent already built or in pilot:
- Week 1. Inventory every agent in production or pilot. Define what "correct" means for each, in writing.
- Week 2. Build a baseline evaluation suite: twenty to thirty personas covering your highest-volume intents, plus an adversarial set.
- Week 3. Run the baseline. Expect the first results to be uncomfortable, particularly on context retention and Arabic dialect handling.
- Week 4. Set thresholds, wire the suite into CI, and make the gate blocking for the dimensions you cannot afford to fail.
Teams still at the design stage should build the evaluation suite before the agent. The suite is the specification.
Frequently asked questions
What is agent-to-agent testing?
Agent-to-agent testing is an approach where autonomous evaluator agents interact with an AI agent in a controlled environment to assess how it understands, responds and collaborates. Responses are scored across dimensions such as accuracy, hallucination, bias, toxicity and context awareness rather than marked simply pass or fail.
How is it different from normal test automation?
Conventional automation checks deterministic outputs against fixed assertions. AI agents produce different responses to identical inputs, so they must be evaluated statistically across many simulated conversations rather than asserted against a single expected result.
Is TestMu AI the same company as LambdaTest?
Yes. LambdaTest rebranded to TestMu AI on 12 January 2026. The team, infrastructure, customer accounts, integrations and pricing structure remained the same, and existing credentials continue to work.
Can it test voice agents and Arabic-language agents?
Yes. The platform supports evaluation of inbound and outbound voice callers using more than two hundred voice profiles, covering accents, interruptions and multi-language conversations including Arabic.
Does it replace our QA engineers?
No. It removes repetitive regression and maintenance work so QA engineers can focus on exploratory testing, risk assessment and defining what correct behaviour means. Humans still set the thresholds and own the release decision.
Do we need to abandon our existing test scripts?
No. Existing Selenium, Cypress, Playwright and Appium suites run on the grid without rewriting, and newly authored tests export back into those same frameworks.
How does this connect to Google's A2A protocol?
The A2A protocol standardises how agents from different vendors discover and delegate tasks to each other. Agent-to-agent testing validates whether those agents behave correctly and safely once they do. Organisations deploying multi-agent systems generally need both.
Where does Pro Skills deliver this?
Across Saudi Arabia, from our Riyadh base, with delivery extending across the wider EMEA region.
Pro Skills Trainings & Consulting