What CodeSignal Records When AI Is Allowed

CodeSignal now runs two opposite regimes on the same platform. In a standard proctored assessment, generating code with an outside AI tool is the thing the integrity checks look for. In an AI-assisted or agentic assessment, using AI is the task — and every prompt you type is recorded, labelled, and handed to the employer alongside your code. Most candidates cannot tell from the invitation email which regime they are about to enter, and the two reward exactly opposite behaviour.

This guide covers the AI-permitted side: what CodeSignal captures when an assistant is switched on, who reads the transcript, what the evaluation is actually measuring, and how to tell which kind of test you have been sent.

The two products that allow AI

CodeSignal has shipped AI-permitted assessment twice, and the difference matters.

Cosmo, the in-IDE assistant. CodeSignal announced AI-assisted coding assessments and interviews on 30 May 2025, built around an assistant called Cosmo "embedded directly into our state-of-the-art integrated development environment (IDE)." The company describes two configurations an employer can choose: "Full AI Co-Pilot Mode" for active collaboration, and "Guided Support Mode" for "light-touch assistance for navigation and syntax." Cosmo lives inside CodeSignal's own editor; it is the assistant the test hands you.

Agentic coding assessments. On 2 April 2026 CodeSignal announced agentic coding assessments, a different shape. The release describes a candidate task in three parts: "extract and interpret product or technical requirements," "use agentic AI tools such as Claude Code, Cursor, and Codex to build a working solution," and "explain their technical decisions and reasoning to a human reviewer." Co-founder and CEO Tigran Sloyan framed the rationale as "engineers are no longer coding alone; they're working with AI agents, and the best ones know how to get the most out of them."

CodeSignal's agentic assessments page describes the task types as "AI chat interactions, prompt design challenges, AI-assisted coding exercises, and scenario-based problem solving," says employers get "flexible LLM support" to "select which large language models power your assessments," and puts length at "20–30 minutes" for AI literacy assessments and "60–90 minutes" for "more advanced technical assessments." As of 15 September 2026, CodeSignal's homepage headline is "Agentic skills validation & development," and Agentic Assessments sits in the Assess menu rather than as a front-page product.

What actually gets recorded

The prompt transcript is the artifact, and CodeSignal's own documentation is blunt about it.

Its knowledge base article on evaluating AI skills with Cosmo states that "all of a test-taker's interactions with Cosmo are logged and made available to you," and tells reviewers exactly where to look: "from the test-taker's coding report. Click into any question and scroll down to the area labeled AI Assistance Transcript." For live interviews, the AI Co-Pilot for Interview documentation says the "full conversation is logged in the interview coding replay," reachable by clicking a Cosmo icon inside the replay.

Two things follow that candidates routinely underestimate.

First, the transcript is not a summary. It is the conversation, turn by turn, sitting next to the session replay that already captures your keystrokes and edits. A reviewer scrubbing your replay can see the moment you asked for something and the moment the answer landed in your file.

Second, CodeSignal states the purpose plainly. The documentation says the log gives "context for whether the test-taker used Cosmo to generate code, ask about the environment, or something else." That distinction — did you outsource the problem or clarify the environment — is the thing a human is reading for.

You are told this before it starts. CodeSignal's documentation says that "before a test-taker begins to interact with Cosmo, they are informed the conversation will be recorded and viewable by both CodeSignal and your company." Treat that notice as the signal it is: the prompts are part of your submission.

How to tell which regime you are in

The invitation email rarely says. The reliable tells are in the environment itself.

• An AI panel is present in the IDE. If Cosmo or an equivalent assistant is available in the editor, AI use is enabled and logged. Employers must have it turned on at the account level — CodeSignal's documentation routes organisations through their Customer Success Manager before individual assessments can toggle it — so its presence is deliberate, not an oversight.
• A consent notice mentions AI recording. The pre-interaction notice about the conversation being recorded and viewable is specific to AI-enabled sessions.
• The task asks you to interpret requirements rather than solve a stated problem. The agentic format leads with extracting requirements and ends with explaining your reasoning to a person. A classic algorithmic prompt with hidden test cases is the older format.
• No assistant, plus a setup flow for camera, microphone, and ID. That is a proctored session, and the rules are the opposite ones. How to tell whether a CodeSignal test is proctored before you start covers the setup steps in order.

If an assistant is not offered and the instructions do not permit outside tools, assume AI is not allowed. Do not infer permission from the fact that CodeSignal sells AI-assisted products.

What the scoring is measuring

CodeSignal has not published the rubric. Its agentic assessments page says results "show proficiency levels across specific AI competencies," describes the assessments as "validated by I-O Psychologists" and "informed by over 13 million skill evaluations," and stops there — no competency list, no weights, no thresholds. The April 2026 release is similarly high-level, saying the assessments evaluate "how candidates produce real work with AI tools." Take the psychometric framing as a claim about process, not as a published scoring key.

What the documented artifacts imply is narrower and more useful. If the transcript, the replay, and a human explanation round are the evidence, then the evaluation has room to weigh how you decomposed the problem, whether you checked what the agent produced, and whether you can account for code that carries your name. A solution that works but that you cannot explain is a worse outcome here than in a traditional assessment, because the explanation round is built into the format.

One practical consequence: narrate inside your prompts. Asking an agent to "add input validation for the empty-list case and explain the tradeoff" leaves a different record than "fix this." The first reads as direction, the second as delegation, and the transcript is what a reviewer has.

The open question about the Suspicion Score

CodeSignal's standard integrity machinery includes an automated Suspicion Score built on editor activity, with pasted code among the signals it weighs, plus human review of proctoring recordings. We cover both in how CodeSignal detects cheating and what the Suspicion Score means.

In AI-assisted mode, code arriving in the editor from an assistant is the expected behaviour rather than an anomaly. CodeSignal's public documentation does not state how, or whether, the Suspicion Score is adjusted when Cosmo is enabled, and we found no primary source addressing it as of 15 September 2026. That is a genuine gap, not a subtlety we are glossing: the platform documents the transcript thoroughly and the interaction between AI permission and automated integrity scoring not at all.

The safe reading is that permission is scoped to the assistant you were given. Using the provided in-IDE assistant is the documented, logged, intended path. Routing work through an outside tool produces no transcript, which in a format where the transcript is the evidence leaves a conspicuous hole rather than a clean record. Do not treat "AI is allowed" as a general waiver.

Where this sits in the wider shift

CodeSignal is not alone. Several large employers and platforms now run interviews where AI is permitted and the prompt history is reviewable, which changes preparation more than it changes the work.

It is also worth separating this from CodeSignal's other AI product. An AI-assisted assessment is you using AI while humans review the result. An AI-run interview is software conducting the interview and scoring you; what CodeSignal's AI Interviewer records and scores covers that one. The invitation wording for the two can look similar and the preparation is not the same.

What to do with this

Assume the transcript will be read by a person who also has your replay. Prompt in complete thoughts rather than fragments. Review what the agent gives you before it goes in, and say in the prompt when you are rejecting something and why — that is the cheapest evidence of judgment you can leave. Expect to explain the result out loud.

Control is a desktop AI interview assistant, and it is the wrong tool for this format. When the assessment supplies its own logged assistant, an outside one adds nothing a reviewer can credit and removes the record the evaluation runs on; in the explanation round that closes an agentic assessment, you will be asked about decisions the transcript shows you did or did not make. The preparation that pays here is practising the explanation itself — our guide to running an AI mock interview with a scoring rubric is a closer fit than any live assistance.