Constraints over conversations
If a rule matters, make it machine-checkable.
THE ADD FRAMEWORK / 01
A proposed operating model for software projects where agents do the building and humans govern intent, constraints, and exceptions.
THE PREMISE
ADD preserves the purpose of coordination and quality, then expresses it through specifications and checks a machine can execute. Human collaboration still matters; delivery no longer has to wait for a sprint boundary.
| Scrum artifact | ADD replacement | Purpose |
|---|---|---|
| Sprint | Continuous flow | Work advances when its gates pass. |
| Backlog | Spec queue | Versioned intent, constraints, and acceptance criteria. |
| Story points | Budget & actual cost | Tokens, compute, and human time. |
| Stand-up | Live telemetry | Failures and blocked tasks surface as exceptions. |
| Status column | Gate state | Evidence from the pipeline determines progress. |
| Sprint board | Execution graph | Tasks, executors, dependencies, cost, and proof. |
EIGHT PRINCIPLES
If a rule matters, make it machine-checkable.
Humans own specifications. Code is an output.
Acceptance criteria become tests before execution.
Use the cheapest tier that meets the quality bar.
Humans govern policy and handle exceptions.
Least privilege; no secrets in prompts.
Link changes to their spec, executor, and cost.
Use failures to improve constraints and routing.
4 SPEC ARTIFACTS
The stack, dependencies, budgets, and data rules agents must obey.
Versioned ADRs explain why the system is shaped the way it is.
Given / When / Then. Every criterion maps to a test before execution.
Required coverage, test types, API documentation, and changelog updates.
Readiness rule: every acceptance criterion maps to a test. Ambiguity returns to a human before execution.
6 QUALITY GATES
Formatting, lint, type checks, and constraint policies.
Secret scanning, dependency audits, SAST, and prompt-injection checks.
ADR compliance, module boundaries, and forbidden dependencies.
Acceptance tests and coverage thresholds.
Required docs, API references, and changelog changes.
Preview smoke tests, performance budgets, and rollout checks.
Agents open pull requests. They do not merge directly, modify pipelines, or receive secrets in prompts.
4 + 1 EXECUTION TIERS
| Tier | Executor | Typical work |
|---|---|---|
| 0 | No model | Scaffolding, code generation, formatting |
| 1 | Small model | Boilerplate, CRUD, docs |
| 2 | Mid model | Feature logic, single-module refactors |
| 3 | Frontier model | Cross-module design, security code |
| Last stop | Human | Ambiguous specs and exhausted escalations |
Framework policy: retry once with the gate error, then move up a tier. Use passing history to improve routing.
MEASURE THE OUTCOME
Tokens, compute, and human minutes required to ship.
Tasks passing every gate without a retry, by tier.
Tasks moving up a tier or reaching a human.
Time from ready specification to production.
Ambiguous specs returned before execution.
Deployments requiring rollback or a hotfix.
7 REPLACEMENT CEREMONIES
Every ceremony starts with evidence produced by the platform and ends by changing an artifact. If no spec, constraint, test, ADR, routing rule, permission, or priority changes, the meeting was unnecessary.
Assign every failed task or ambiguous spec. Cancel when the inbox is empty.
Replaces the daily stand-upReject ambiguity before tokens are spent. Every criterion must map to a test.
Replaces refinement and planningTurn repeated gate failures into constraints, tests, templates, or ADR proposals.
Replaces the system part of retroCompare cost, first-pass rate, and escalation patterns; update tiers and budgets.
Replaces velocity trackingApprove ADRs and the architecture tests that make each decision enforceable.
Replaces ad-hoc design reviewCompare shipped behavior with the need, then reprioritize the spec queue.
Replaces sprint reviewRemove unused access, inspect anomalies, and tighten agent policy by default.
Adds the review most teams miss4-STAGE ADOPTION PATH
Make the Definition of Done executable for human and agent work.
Choose a low-risk area and run approved specs through agent pull requests.
Add execution tiers, budgets, the failure review, and an exception rotation.
Move work to the spec queue and replace sprint ceremonies with ADD reviews.
RISK IS MOVED, NOT REMOVED
Human-approved acceptance criteria and mutation testing protect against trivial tests.
Outcome reviews tie delivered specs back to product behavior and stakeholder feedback.
Sandboxing, least privilege, scoped credentials, and security checks limit the damage.
Per-task and per-spec budgets stop loops and surface anomalies for review.
Every constraint links to its failure or ADR and is pruned when it no longer earns its cost.
Exception rotation, ADR ownership, and deliberate pairing preserve technical knowledge.
ABOUT THE FRAMEWORK
Based on Agent Driven Development: Managing a Team That Builds With AI, published 24 October 2026. RustGrid is the reference implementation proposed by the paper. The framework defines the operating model; the product and roadmap show which parts exist now and which complete the control plane next.
THE NEXT WAY TO BUILD
Start with RustGrid’s execution foundation. Build toward a delivery system governed by specs, gates, and clear ownership.