May 24, 2026AI/LLM Updates

Claude Code's "Routines" and "Dreaming" Are About to Change How You Run CI/CD Tests

Why it matters for testing

Anthropic's two new Claude Code features — Routines and Dreaming — shift automated testing from human-triggered scripts to autonomous, self-learning agents that can plan, execute, review, and iterate on tests across the full pull request lifecycle with zero babysitting.

Intro

For years, "AI-assisted testing" meant a developer pasting a function into a chatbot and getting a handful of unit tests back. Useful, but not transformative. What Anthropic announced in May 2026 is a different order of magnitude: Claude Code can now schedule itself, respond to GitHub events, track CI failures over the life of a pull request, and — critically — teach future instances of itself what it learned along the way. If you run a CI/CD pipeline with meaningful test automation, this changes your workflow.

The AI development/news

Anthropic rolled out two distinct but complementary Claude Code features in May 2026.

Routines allow developers to configure automated Claude Code workflows that trigger in three ways: on a schedule (cron-style), via an HTTP endpoint, or in response to GitHub webhook events. A webhook-based routine, for example, can watch for pull requests that match specific branch patterns, spin up a Claude Code session automatically, monitor CI failures as they come in, respond to review comments, and keep working across the full lifecycle of the change — all without a developer manually invoking the tool.

Scheduled routines cover recurring QA housekeeping jobs: triaging stale bug reports, detecting documentation drift against updated code, generating regression test drafts when new commits land. API-triggered routines let external systems — deployment pipelines, monitoring platforms, internal dashboards — call Claude Code sessions via authenticated HTTP requests.

Dreaming tackles a different problem: knowledge retention across agent sessions. When Claude Code agents work on tasks, they now write structured notes to themselves. When another agent later works on the same codebase, it can read those notes to understand prior decisions and known failure modes. Dreaming is the consolidation layer: Claude Code periodically reads all accumulated notes, spots patterns and recurring issues across different tasks, and synthesizes them into durable institutional knowledge. The goal, in Anthropic's words, is to push automation "as far as it will go" — not just generating code, but checking and correcting its own work using what it learned from previous runs.

Current testing landscape

Today, most CI/CD pipelines invoke test suites at fixed points: on commit, on PR open, on merge. Those test suites are static — they run the same scripts every time, report results, and stop. Human engineers then read the logs, diagnose failures, update tests when the app changes, and repeat the cycle. AI has entered this loop mainly at the generation layer: tools like GitHub Copilot and Claude can write test stubs when asked. But they don't monitor, adapt, or retain knowledge between runs.

The result is that test maintenance is still largely manual work. Tests break when UIs change, when APIs shift, when new dependencies are introduced. Someone has to notice, diagnose, and fix. AI generates; humans maintain.

The impact

Routines close the monitoring gap. A Claude Code workflow triggered by a GitHub PR webhook can watch the PR's CI results in real time, identify which tests failed, hypothesize causes, update test code to match application changes, push fixes, and watch the next CI run — all without a developer stepping in. This is not speculative; it is the explicit use case Anthropic documented in the May 2026 launch.

Dreaming closes the institutional knowledge gap. Today, every CI run is memoryless — the agent that helped fix a flaky test last Tuesday knows nothing when it's invoked again next Tuesday. With Dreaming, patterns across runs accumulate. If a particular module consistently produces timing-related test failures after a specific type of refactor, future agents will know this before they start. Debugging time drops; false-positive flakiness becomes recognizable noise rather than investigation work.

Together, these features begin to describe an autonomous QA loop: detect change → generate/update tests → run → observe failure → diagnose with prior knowledge → fix → re-run. The human role shifts from operating that loop to setting its quality thresholds and reviewing its outputs.

Practical applications

Automated regression triage: Set up a scheduled Routine to run nightly, scan recent commits, and flag any tests that are likely to fail based on what changed. Human engineers arrive to prioritized, pre-diagnosed failures rather than raw red builds.

PR-lifecycle test maintenance: Configure a webhook Routine to trigger when a PR is opened against your main branch. Claude Code monitors CI, responds to test failures by updating test selectors or data fixtures, and leaves a structured comment explaining what it changed and why — traceable, reviewable, not a black box.

Cross-session flakiness memory: Let Dreaming accumulate context over weeks. Ask Claude Code "what have been the most common failure patterns in our payment module tests in the last 30 days?" and get a synthesized answer informed by every previous session that touched that module.

Documentation drift detection: Schedule a weekly Routine to compare your API docs against your integration test suite. Flag endpoints that are tested but undocumented, or documented but not covered.

Tools/frameworks to watch

Claude Code (Anthropic) — The core platform. Routines and Dreaming are live features; the Claude Console on AWS includes prompt evaluation tools for building and testing Claude-powered workflows.
GitHub Actions + Claude Code webhooks — Native integration path for PR-lifecycle automation.
jcode — A specialized open-source framework (trending on GitHub, May 2026) for evaluating code agents' reliability, useful for testing your Claude Code Routines themselves before deploying them to production.
Playwright / Vitest — Still solid test execution layers; Claude Code Routines orchestrate above them, not instead of them.

Conclusion

The gap between "AI helps write tests" and "AI manages the test lifecycle" has been closing for two years. Claude Code's Routines and Dreaming features are a meaningful leap across it. The teams that benefit most in the near term will be those who treat these features as infrastructure — setting up PR-triggered Routines and letting Dreaming accumulate context across real production cycles — rather than one-off experiments. The longer the system runs, the smarter it gets about your specific codebase. That compounding value is the core proposition, and it's worth taking seriously now.