What are cascades?
Cascades is a structured delivery methodology that replaces ceremony with discipline enforced by software. Delivery runs through four phases, each producing an artifact the next phase consumes. The platform enforces the structure while the human provides direction and judgment, so nothing depends on one person holding the methodology in their head. The platform itself was built end to end through the cycle.
The four phases
- Shape. Business conviction becomes structured, testable work. A business analyst agent translates a goal into paired PRD sections and tech spec sections, decomposed into tasks with Given/When/Then acceptance criteria. Dependencies are mapped, and sequencing follows what unlocks the most downstream value.
- Gate. Spec gates evaluate every PRD and tech spec pair on three criteria: integrity (buildable specs, testable acceptance criteria, traceability between requirements and tasks), collisions (contradicting sections, the same pattern specified three different ways, product requirements contaminated with architecture), and consequences (dependency gaps, missing tasks, sections with no owning task).
- Ship. Persistent coding agents execute the shaped, gated work. Each agent reads its full task directive, writes code, runs tests, and posts progress commentary, all logged for audit.
- Verify. Acceptance criteria locked in Shape are verified against shipped code. CI gates, regression tests, and agent code review run, and a human confirms the result before a task moves to done. Failed work routes back to Shape (wrong spec) or Ship (wrong implementation).
Each phase cascades into the next: Shape produces the spec Gate evaluates, Gate produces the clean spec Ship executes, and Ship produces the code Verify tests. A spec contradiction caught in Gate never reaches code, and a failing test never reaches done; the cascade contains problems in the phase that catches them.
Handling parallel agent work
Multiple agents work in parallel on independent tasks, and dependency chains prevent an agent from starting blocked work. Agents persist across sessions, keeping context in place. When a test fails, the agent reads the failing check, re-plans, creates a new attempt, and changes its approach after diagnosing the prior failure. Transient failures (rate limits, timeouts, flaky infrastructure) retry up to three times with exponential backoff. Past the retry budget, the agent stops, posts a comment documenting each attempt and why it failed, and sets the task to blocked. Nothing auto-closes, and failed work is never marked done.
When the cycle pays off
The methodology targets failures where they are cheapest to catch. A failure caught upstream costs hours to fix; the same failure caught downstream costs weeks of rewriting plus rework across dependent tasks. Shaping eliminates waste from bad cross-domain communication and turns a week of cross-functional meetings into about 40 minutes of agent shaping, with testable output.
How the work divides
Agents produce structure and code; humans direct and confirm. In Shape, the human brings the conviction about what matters and the agent brings the structure that makes it shippable. The Gate runs on deterministic automated checks plus agent reasoning, with no human review required. In Ship, humans monitor rather than manage. In Verify, automated validation runs first and human verification comes last, so tasks move to done only after a human confirms.
Cascades is the methodology; AIDLC is how you run it, the five-stage operating cycle a piece of work travels when agents do the building. The gates exist to catch drift, and the catalog of forms drift takes is on the What is drift page.
What happens when a spec gate fails?
Work is blocked, and the spec stays blocked until it is clean. Findings are specific and actionable: they point to the exact section and the exact problem. Gate checks combine deterministic automation with agent reasoning, and no human review is required for the gate to run. A clean gate produces a spec ready to build.
What happens when a test fails during Verify?
The failure is detected automatically, so nobody has to watch CI. The agent reads the failing check, re-plans, and retries with a changed approach. Transient failures are retried up to three times with exponential backoff. Past the retry budget, the agent documents each attempt and why it failed, then sets the task to blocked for a person to decide. Failed work is never marked done.
What replaces sprints, standups, and sprint planning?
Task commentary is the status report, always current and readable as a decision log, so no meeting is needed for status. Sequencing follows the dependency graph, which updates as work completes instead of timeboxes. The platform enforces the methodology, so no human bandwidth is spent maintaining the process.