Gremlin's Foresight AI agents wait for an engineer before every test and fix

The part worth copying is the loop. When a test exposes a weakness, the Operator agent proposes a change and the service reruns the original test ^[5][6]. The proof that a fix worked is the experiment that failed, now passing. Andrus said: "An alert going quiet doesn't prove anything. A test that passes under load, in realistic conditions, does." ^[8] We think rerunning the failed test is the right standard. A quiet alert shows a symptom stopped. A passing fault injection shows the failure mode is handled under the conditions that test models.

The fourth agent, a Technical Program Manager, tracks test coverage, reliability risks and commitments across teams, then writes the weekly report and posts Slack updates ^[5]. Few engineers will miss doing that job. The service also builds dashboards from natural-language requests, meant to show changes in reliability scores and the risks teams have addressed ^[9].

The second sound decision is where the language model sits. According to Gremlin, the agents decide from platform data and the Failure Atlas, and the LLM handles search, summaries and explanations ^[16]. That keeps a text generator out of the step that chooses which fault to inject. Gremlin also says system data is not used to train a model ^[16].

Treat the Failure Atlas like a benchmark table. Gremlin describes it as a proprietary record of millions of fault-injection experiments run over more than ten years across tens of thousands of distributed systems ^[15]. Its recommendations transfer to a service only if that service fails the way the corpus did, with comparable dependency graphs and comparable retry and timeout behaviour. Because the record is proprietary, a buyer can only test the fit by running it against their own services ^[15].

Gremlin says Foresight AI will neither run a test nor apply a remediation without approval ^[7]. It positions the product against AI-assisted delivery, with Andrus arguing that more code is reaching production without close human review ^[4]. In InfoQ's account he does not attach a figure to that ^[4]. So a product aimed at a review shortfall still asks a person to approve every test and every fix. The engineer moves from writing experiments to reviewing them. We'd expect a review to cost less than authoring one, but the approval queue still grows with each service enrolled.

Wix has described using AI in CI/CD for log interpretation and remediation suggestions, with deployment and critical infrastructure decisions left to humans ^[10]. Gremlin's gate covers running the test as well as applying the fix, and we think it belongs there for a first GA release of a tool whose job is injecting failures ^[7]. Cloudflare's Agent Development Lifecycle proposal has agents managing more of testing, deployment and maintenance ^[11].

Gremlin's own claim about expertise is narrow. The product is meant to provide reliability checks without every developer becoming a distributed-systems specialist ^[3]. Engineer Sandip Bhattacharya argued in a LinkedIn discussion that organisations still need specialists to configure agents ^[12]. Intellyx analyst Jason English, in an article written for Gremlin, cautioned that a reliability programme cannot simply be installed as a software package and needs shared ownership and accountability across engineering and management ^[13]. Intellyx discloses that Gremlin is its customer ^[14].