Wave · Self-Adaptive V&V Orchestration
Wave classifies tasks, selects pipelines, validates outputs, and learns from results. Contract-validated. Gate-controlled. Self-correcting.
Just the right amount of guardrails.
Agent factories need boundaries — not to hobble agents, but to make them trustworthy enough to run unsupervised.
Approval loops at every step. Agents ask before breathing. Safe on paper — useless in practice. You're still doing the work.
Each persona is fully empowered inside its role, hard-constrained outside it. Scoping is declarative, enforced at runtime, and versioned in git.
Unconstrained agents with full codebase access. One misread prompt from leaked secrets, deleted files, or broken code in production.
Self-Adaptive Orchestration
Describe a task and Wave classifies it, selects the right pipeline, executes it, and records the outcome. The feedback loop improves routing over time.
Pipelines as Code
Define multi-step AI workflows in YAML. Version control them, share them, run them anywhere.
Contract Validation
Every step validates its output against schemas. Get structured, predictable results every time.
Git-Native Workspaces
Steps execute in real git worktrees right inside your repo. No detached folders, no mount hacks — just native git.
Audit Logging
Complete execution traces with credential scrubbing. Full visibility into every pipeline run.
Ready-to-Run Pipelines
82 built-in pipelines for code review, security scanning, documentation, issue implementation, and orchestration — ready to plug into your agent factory.
Human & Automated Gates
Four gate types — approval, timer, pr_merge, ci_pass — let you pause pipelines for human review, wait on timers, or poll for PR merges and CI checks before proceeding.
Rework Loops & Convergence
When a contract fails, Wave feeds the error back and retries. Convergence tracking detects score plateaus and aborts stalled loops to save tokens.
Project Ontology
Declare bounded contexts, invariants, and conventions in wave.yaml. Wave injects domain knowledge into steps and tracks context success rates. Behind feature flag.
Pipeline Composition
Five composition primitives — sub-pipelines, iterate, branch, loop, and aggregate — let pipelines compose other pipelines for complex multi-stage workflows.
Automatic Model Routing
A 3-tier routing system — cheapest, balanced, strongest — classifies step complexity by persona and composition usage, routing each step to the right model tier automatically.
Meta-Pipelines
A philosopher persona dynamically generates and executes child pipelines at runtime, with configurable depth (3), step (20), and token (500K) limits.
Web Dashboard
Monitor pipeline runs, visualize step DAGs, browse artifacts, and control execution — with real-time SSE updates and token-based remote auth.
How Wave Verifies Agent Work
A four-layer verification & validation model ensures every pipeline output meets quality, structural, and behavioral requirements.
Bounded contexts inject domain invariants and conventions into agent sessions. Lineage tracking measures which contexts produce the best outcomes. Behind feature flag — not shipped by default.
Role-scoped agents with controlled tool access, temperature, and git forensics capabilities for each pipeline step.
11 contract types with rework loops and convergence tracking. Self-correcting steps retry with feedback until quality thresholds are met.
Four checkpoint types — human approval, timed waits, PR merge polling, and CI pass polling — with toast notifications for attention states.
Any agent. Any forge. One orchestrator.
Wave speaks your agent's language and works with your platform from day one. No lock-in, no migration.
Switch adapters at runtime with --adapter and --model. Mix them in a single pipeline.
From the re:cinq blog
"The factory sets boundaries on what's safe to do, not what's allowed."Building Agent Factories →