AI coding agents: OutSystems opens its platform to Claude Code, Cursor and Codex — and writes the code itself
OutSystems has made its Agent Experience generally available, opening its enterprise platform to external AI coding agents — Claude Code, Cursor, Codex and Kiro. The move, announced at the company’s World Tour event in Las Vegas, inverts the usual arrangement. Instead of letting an agent write code freely, OutSystems lets the agent work at the level of system design, then generates the underlying code itself, deterministically, with security, testing and governance controls applied.
It is a striking position for a company whose business is a low-code platform — a system built on the premise that you assemble applications inside it, not alongside it. By inviting the AI coding agents that threatened to bypass it, OutSystems is betting that the answer to messy AI-generated code is not a better model but a governance layer.
What OutSystems announced
Agent Experience opens the OutSystems platform and its Enterprise Context Graph — a live model of a customer’s systems, data and logic — to third-party coding agents. A developer describes what to build using a preferred agent; the agent shapes the architecture; and a platform-native agent called Mentor produces the same application model as OutSystems’ own development studio. The connection runs through OutSystems MCP, an implementation of the Model Context Protocol that the company says is generally available on every OutSystems Developer Cloud tenant.
Connecting through MCP matters because it is becoming a common interface: rather than building a bespoke integration for each AI coding agent, OutSystems exposes its platform once and lets any compliant agent use it. That lowers the cost of supporting new agents as they appear, and it keeps the platform’s governance in one place. It also means OutSystems’ reach now depends on agents it does not control — a dependency the company has chosen over competition.
Build With Any Agent
The platform bundles three promises: build with any agent while keeping architectural integrity; stay in control through built-in security, automated testing and governance; and deploy anywhere — public cloud, private cloud, on-premises or hybrid — which OutSystems frames as “operational sovereignty” for regulated industries. Alongside Agent Experience, the company announced Agent Workbench, Agent Spotter and industry toolkits, including an insurance offering due in November. For regulated industries, the pitch is straightforward. Banks, insurers, healthcare providers and law firms cannot let agents write freely into production systems, but they want the productivity of AI coding agents.
Agent Experience offers a middle path: use the agents, keep the controls. The same logic explains the industry toolkits — pre-built claims agents for insurers and loan-application components for banks — which give those buyers a starting point rather than a blank platform. The open question is durability: the agents OutSystems now supports are updated frequently and controlled by other companies, so the platform’s governance must keep pace with tools it does not own.
The inversion: agent designs, platform generates
The interesting part is the division of labour. Most AI coding tools generate code directly, and the output is only as good as the model’s grasp of the surrounding system. OutSystems inverts that: the agent operates on the design, and the platform generates code from its model. The company argues this minimises technical debt, maximises code reuse and eliminates “fragmented AI-generated code in production.”
That is a meaningful distinction, because it changes what the agent is trusted with. An agent that writes code needs to understand dependencies, data relationships and business logic on its own; an agent that only shapes the design hands those constraints to a platform that already models them. The catch is that the same design also constrains what the agent can produce — you get governance and coherence, but you build within OutSystems’ model, not anywhere.
The problem it targets
OutSystems is not inventing a problem. The reliability of AI-generated code is one of the more documented issues in enterprise software. New Relic’s 2026 State of AI Coding report, based on a survey of 200 United States technology decision-makers, found that 67% say AI generates or refactors between 51% and 75% of their weekly code output, and that while 94% rate AI code as higher quality at review, 78% report more production incidents and 74% report that at least a quarter of AI code needs significant rework. Lightrun’s 2026 study found that 43% of AI-generated changes need manual debugging in production even after passing QA and staging. Sonar found that 96% of developers distrust whether AI code is functionally correct.
The release cites “recent developer survey data” showing that “74% of AI-generated code requires significant post-generation rework or is never deployed to production.” That figure is close to New Relic’s — but not identical. New Relic reports that 74% of leaders say at least 25% of AI code needs significant rework; OutSystems’ phrasing is broader, and the release does not name the survey. The direction is well supported; the exact number deserves care.
Why “deterministic” is the claim to test
The word “deterministic” carries the argument. In a model-driven platform, code is generated from a defined model rather than improvised by a model, so the same inputs should produce the same structure. That is a real difference from free-form generation, and it is the basis for the governance promise: predictable, reviewable, testable output.
But OutSystems does not define how far the determinism extends, and no independent test yet shows that agent-designed, platform-generated code fares better in production than agent-written code. The claim is plausible — model-driven generation is more constrained than a language model’s free output — and it remains a claim. The customer results that follow are early-access accounts, not audited benchmarks.
The strategic reversal
The move also reads as self-preservation, executed well. AI coding agents let developers generate code outside any platform, which is precisely the threat to a low-code vendor’s model. OutSystems could have competed with them; instead it opened to them and positioned itself one layer up, as the governance and architecture system that makes the agents safe for regulated enterprises. That is a bet that enterprises will pay for control rather than for generation — a plausible read of a market where the generation is becoming commoditised. The commercial logic is clear. Governance, auditability and deployment sovereignty are things enterprises pay for directly, and they are harder to commoditise than code generation.
By contrast, the generation layer is crowded and cheapening fast, with Claude Code, Cursor, Codex, Copilot and Kiro all competing on similar ground. If governance is the durable margin, OutSystems has picked the right layer; if developers simply route around the platform, it has handed the relationship to its rivals.
Who OutSystems is
OutSystems, founded in 2001, sells an enterprise application platform and says it has more than 90 million end users, over 600 partners and thousands of customers in more than 75 countries across 20-plus industries, with recognition from Gartner, IDC and Forrester and a top ranking for customer satisfaction on G2. It calls itself an “Agentic Systems Platform,” a repositioning that Agent Experience is meant to substantiate.
The customer evidence
The release names three early-access customers. YESCO, a sign manufacturer, used Claude Code with the platform to compress software release cycles to two weeks. Lowenstein Sandler, a United States law firm, built a production-ready litigation matter tracker in under three hours, cutting application finishing times from an hour to five to ten minutes. Normal, a Danish retailer, had a single developer build an application in two weeks, down from about six. Each is from a company result from a customer with an interest in the programme with no is independent audit.
What to watch
For enterprise buyers, the questions are pricing, the limits of the governance model, and how much of a project must live inside OutSystems to get the promised coherence. For OutSystems, the test is whether opening to external AI coding agents grows the platform’s reach or simply makes it a back end for agents that own the developer relationship. Moreover, or the market, Agent Experience is an early example of a platform betting that governance, not generation, is where the durable margin sits. And, for platform teams, the practical test is whether governance travels with the code. If an agent builds inside OutSystems, the controls hold; if the same AI coding agents are in use elsewhere in the estate, the coherence stops at the platform’s edge. That boundary is where the promise is most likely to fray.
What the announcement does and doesn’t establish
It establishes a generally available product, a defined set of supported agents and a clear architectural idea. It does not establish that deterministic generation outperforms agent-written code in production, that the governance promise holds across complex estates, or that the cited rework figure is the company’s own. Moreover, t is a well-argued answer to a real problem — and, for now, the argument is OutSystems‘.
The honest summary is that OutSystems has made a smart, counter-intuitive bet: that the way to keep developers in an agent-driven world is to let the agents in, and sell the guardrails. Whether that bet pays will show in what enterprises actually build — and whether they build it inside the platform or merely through it. That is the bet to watch.
Editor’s Note: This article draws on the company’s announcement and public product materials, plus independent press coverage and independent research on AI-generated code quality, including the New Relic 2026 State of AI Coding report, Lightrun’s 2026 study and Sonar’s developer survey. The customer results are from the company-report early-access accounts and have no independent verification, and the release’s cited “74%” figure does not name its source. No interview. See the source list in the publication package.