Guided Intelligence
Read paper
Whitepaper · An AI-native engineering operating model

Guided Intelligence

A Human-in-the-Loop, AI-Native engineering operating model built for 10× velocity without losing stability. Humans set direction. AI does the work. The pipeline keeps moving.

Read the whitepaper → Explore concepts See the flow
Roles
Planner · Builder · Reviewer
Flow
Continuous · single-threaded
Safety
7 layers · Sev1–Sev5
Geometry
Domain × Track × Cell
§1 · The problem

The constraint moved. The process didn't.

For two decades, productivity gains came from layering on top of Agile and DevOps. AI breaks that equilibrium and exposes the mismatch between what teams can now do and how they're organized to do it.

01

Agile breaks under AI acceleration

Sprint cycles, QA phases, and staged review gates were built to stabilize work moving at human speed. When execution takes seconds, those same stabilizers turn into bottlenecks.

02

Coordination is the new constraint

AI absorbs the mechanical work. What is left is the human work around code: clarifying intent, aligning on architecture, reviewing for meaning. Execution gets cheap. Misalignment gets expensive.

03

10× via semantic clarity, not raw output

AI already writes code 100× faster. The 10× system-level multiplier comes from removing the human overhead that surrounds that code, not from typing faster.

Agile
Optimized for human execution
Guided Intelligence
Optimized for AI execution + human semantics
Batch-driven sprint cycles
Continuous unbatched flow
Standups, grooming, planning meetings
Semantic clarity via Execution Plans
Manual user-story decomposition
AI generates scaffolding from intent
PR queues stalled on review
Single Reviewer as semantic gate
End-of-sprint QA batches
PAT validates the running system
Scales with headcount
Scales with AI capability
§3 · Roles

Three roles, plus continuous validation.

GI organizes work around the Planner, Builder, and Reviewer, with continuous PAT validation running alongside. These are responsibilities, not job titles.

All roles →
Planner Architect

Semantic owner of the domain.

Translates ambiguous intent into Execution Plans, constraints, and invariants. Sets the semantic frame the system operates within.

Does
  • →Express intent with crisp boundaries
  • →Define invariants & integration surfaces
  • →Resolve cross-domain contracts
Does not
  • ×Granular task breakdown
  • ×Write scaffolding or boilerplate
  • ×Dictate mechanical sequence
Builder Executor

Flow driver. Supervises AI execution.

Drives AI generation of code, tests, and migrations aligned to the plan. Resolves ambiguity early and keeps the pipeline unblocked.

Does
  • →Drive AI generation aligned to plan
  • →First-pass validation before review
  • →Ask Planner when ambiguous
Does not
  • ×Make architectural decisions
  • ×Expand scope
  • ×Perform semantic approval
Reviewer Semantic Gatekeeper

Final semantic authority before release.

Protects intent alignment, invariant integrity, and domain coherence. The only human at the merge boundary.

Does
  • →Validate intent & invariants
  • →Approve / reject / clarify
  • →Oversee ring promotion
Does not
  • ×Rewrite code manually
  • ×Mechanical / syntax review
  • ×Re-plan or expand scope
PAT Continuous Assurance

Async system-level validation.

Operates outside the Planner-Builder-Reviewer loop. Probes edge cases, runs regressions, and surfaces severity signals from production rings.

Does
  • →Exploratory testing across rings
  • →Maintain evolving regression packs
  • →Surface Sev1–Sev5 signals
Does not
  • ×Gate releases or batch reviews
  • ×Replace Reviewer judgment
  • ×Stop flow except via Sev5
§4 · The Flow

From intent to production, continuously.

One Execution Plan at a time, with no sprint boundaries or release windows. PAT validates the running system asynchronously in the background.

See backflow paths →
PLANNER Execution Plan Scope · invariants BUILDER AI Execution Code · tests · diffs REVIEWER Semantic Gate Intent · invariants DEPLOY Ring 0 → Ring 1 Canary deploy PAT Continuous Async validation INTENT → continuous, single-threaded forward flow → PRODUCTION
§7 · Geometry

Domains, Tracks, and Cells.

A Planner owns a full Domain. Builders and Reviewers certify along one-dimensional Tracks. Execution Plans land in atomic Cells at the intersections.

Explore the geometry →
A GI Domain
Capability Tracks (vertical) × Layer Tracks (horizontal) = Cells
active cell planned cell
Planner
spans the domain
Layer
Mobile UI
Layer
Cloud API
Layer
Data
Layer
Firmware
Capability
Checkout
cell
cell
cell ●
cell
Capability
Identity
cell
cell ●
cell
cell ◌
Capability
Mapping
cell ●
cell
cell ◌
cell
Capability
Notifications
cell
cell ●
cell
cell
§6 · Safety

Continuous validation, layered defenses.

A layered system replaces sprint-based QA with always-on validation. Only a Sev5 stops the line.

All seven layers →
Severity governs flow control
Only Sev5 stops the line. Everything else flows through with appropriate response.
Sev 5
Stop-the-Line
Action Freeze merges & promotions · rollback or flag-off · fix immediately · resume once invariants are restored
Sev 4
Major
Action Flow continues · prioritized fix · optional domain-level constraints
Sev 3
Moderate
Action Flow continues · scheduled fix · extra checks applied if needed
Sev 2
Minor
Action Logged · scheduled · no effect on flow
Sev 1
Trivial
Action Logged · scheduled · no effect on flow
A new equilibrium

Speed with stability, automation with judgment.

GI isn't a radical reinvention. It's the minimum set of structural rules needed for AI-native engineering to work at scale without breaking things.

Read the full whitepaper How to adopt it