Kolton Andrus saw the problem coming. AI tools now push code into production at speeds once unimaginable. The former Amazon and Netflix reliability engineer knew traditional observability and post-incident response would fall short. So his company, Gremlin, built something different.
On October 7, the chaos engineering pioneer made Foresight AI generally available. The agentic add-on doesn’t wait for alerts. It hunts for weaknesses, suggests fixes, runs controlled experiments to test them, and confirms the repairs hold. All of it requires human sign-off at critical steps. Gremlin’s announcement frames the tool as the answer to 10X code velocity creating 10X failure opportunities.
“AI-driven development means shipping code at 10X velocity; it also means 10x the opportunity for bugs, risks, and failures,” Andrus said in the company’s press release. “While the rise of AI SRE tools are great for helping teams respond to incidents faster, it’s still cleanup after something breaks. Gremlin Foresight AI finds risks proactively, delivers the fix, and verifies the fix worked by rerunning the test. This provides teams with the assurances needed to move confidently.”
The system rests on four specialized agents. An Analyst studies service topology, dependencies, and current test coverage to surface likely failure points. A Tester schedules and executes safe fault-injection experiments with blast-radius controls. When tests fail, an Operator examines results against real patterns, then proposes concrete remediations ranging from configuration tweaks to infrastructure code changes. Finally, a Technical Program Manager tracks overall reliability scores, maintains a backlog of commitments, generates weekly narrative reports, and pushes Slack updates to keep teams aligned.
These agents don’t operate on guesswork. They draw from Gremlin’s Failure Atlas. The proprietary knowledge base compiles cause-and-effect data from millions of fault-injection experiments conducted over more than a decade across tens of thousands of distributed systems at large enterprises. InfoQ’s coverage notes this foundation lets recommendations stem from observed outcomes rather than generic advice or speculative root-cause analysis common in other AI operations tools.
Andrus and his team emphasize the distinction. Many emerging AI SRE products focus on faster detection and recovery. Foresight AI instead closes the loop on prevention. It identifies a risk. It recommends action. Humans approve. Tests run again under realistic conditions. Only when the same failure mode no longer breaks the service does the system mark the issue resolved. That verification step, often skipped by busy teams, turns hypothesis into evidence.
Reliability scores give the whole process teeth. Every service receives a measurable rating based on test coverage, detected risks, and validation history. Leaders gain visibility into where investment yields the highest return. Teams see clear priorities instead of vague warnings. The scores update continuously as systems change and new code deploys.
Gremlin built safeguards into the product. Large language models handle summarization, explanation, and natural language chat. The heavy analytical work stays grounded in the company’s platform data and the Failure Atlas. Customer system details never train external models. And every test execution, every proposed change, waits for explicit human approval. The agents augment expertise. They don’t replace judgment.
Beta users already saw the difference. The technology identified risks before incidents occurred and guided teams through validation. One outcome stood out. Fixes that passed repeated targeted tests proved more durable than those confirmed only by the absence of alerts. Production environments stayed steadier even as AI-assisted delivery accelerated change.
This matters now. Organizations adopt generative coding assistants and autonomous agents that generate pull requests with minimal oversight. Release cycles shrink from days to hours. Distributed systems grow more complex with microservices, multi-cloud architectures, and third-party dependencies. Traditional chaos engineering required dedicated specialists and time that many teams lack. Foresight AI aims to distill a decade of hard-won operational knowledge into accessible, repeatable actions.
SD Times captured the shift in perspective. “Gremlin Foresight AI finds risks proactively, delivers the fix, and verifies the fix worked by rerunning the test,” the publication reported, quoting Andrus directly. The story highlighted how the product builds on Gremlin’s original platform for planning experiments, limiting impact, and confirming lasting improvements through automated testing.
Recent discussions on X reinforce the interest. Engineers note that the re-testing of fixes after remediation addresses a common gap. “A fix you haven’t re-broken on purpose is a hypothesis, not a fix,” one observer posted. Gremlin’s own account shared conversations with industry analysts examining the limits of general-purpose LLMs for reliability work and the value of domain-specific data like the Failure Atlas.
The launch arrives at a moment when reliability engineering itself evolves. Site reliability engineering practices brought discipline to operations. Chaos engineering added proactive testing. Now agentic systems promise to scale that discipline across faster development cycles without demanding every developer become a distributed-systems expert.
Gremlin positions Foresight AI as an add-on to its existing reliability management platform. Customers already running fault-injection experiments and tracking service health can activate the agents without starting from scratch. The company, founded by engineers who shaped early chaos practices at Netflix and Amazon, continues to focus on enterprise-scale deployments where downtime carries heavy costs.
Yet questions remain. How well will the agents adapt to novel architectures or emerging failure modes outside the Failure Atlas? Can organizations maintain the necessary human oversight as agent activity scales? Will reliability scores become standardized benchmarks across the industry or remain proprietary signals?
Andrus has signaled confidence. In interviews tied to the launch, he described the product as realizing a long-held vision. Simplify complexity. Reduce toil. Give teams evidence-based confidence to ship fast. The Failure Atlas, built from real production battles, supplies the credibility.
Early reaction from the reliability community mixes optimism with caution. Practitioners welcome tools that surface hidden single points of failure, missing redundancy, or inadequate resilience patterns before customers notice. They also stress that agentic systems work best when integrated into existing observability, deployment, and incident processes rather than operating in isolation.
Foresight AI won’t eliminate all outages. Complex systems always hold surprises. But by shifting reliability work earlier, grounding it in extensive empirical data, and enforcing closed-loop validation, the approach seeks to make failures rarer and less severe. In an era when AI writes more code than humans review, that shift carries weight.
Organizations evaluating the technology should examine how its agents interact with their specific service catalogs, testing schedules, and approval workflows. The natural language chat interface, reliability intelligence summaries, and TPM reporting features may lower the barrier for teams without deep chaos engineering experience. The true test, as always, will come in production under relentless change.
Gremlin made Foresight AI available immediately to existing customers and trial users. The company invites demonstrations that show the system running against its own infrastructure, reportedly achieving high availability targets with the product’s assistance. For reliability leaders balancing velocity against stability, the timing feels deliberate. The agents have arrived just as the pace of AI-assisted delivery threatens to outrun conventional safeguards.
Gremlin Deploys Agentic AI to Head Off Failures Before AI Code Floods Production first appeared on Web and IT News.