Manu's Vault

Intent-Centric-Architecture

4 min readStatus:Seedling

Intent-Centric Architecture is a design approach that treats specifications as the primary durable artifact of a software system — the causal source from which code is generated, not documentation that describes it after the fact.

Core Idea

  • In traditional development, code is the system; in intent-centric systems, intent is the system
  • Code becomes a regenerable output — consumed, discarded, and recreated as requirements evolve
  • This shift is enabled by AI code generation reaching sufficient quality that regeneration is economically cheaper than comprehension
  • The foundational structural element is the content-addressed intent graph: a directed acyclic graph (DAG) encoding causality

Intent graph node types:

  • Requirements and constraints
  • Strategic plans and rejected alternatives
  • Decisions with their rationale
  • Environmental factors
  • Generator metadata (model, template version)
  • Evaluation artifacts (tests, performance budgets, success criteria)

Edges express causality — “this plan depends on that requirement” — not mere association. Each node carries a stable, hash-derived identity (content-addressed), so changing any input changes the node’s address, analogous to how Git hashes work for code objects but applied to the specification tree.

The Versioning Shift

Traditional version control tracks what changed in artifacts (file diffs, line numbers). Intent-centric systems track why a change was required.

  • A single requirement node changes (e.g., email validation must support internationalized domain names)
  • Dependent plan nodes adapt as graph dependencies update
  • Code regenerates downstream as a consequence — no human traces which lines changed
  • The “version” of a component becomes the root hash of the intent graph that produced it
  • Regenerating from identical inputs yields the same identity, making builds reproducible by construction

Why Traditional Traceability Failed

  • Requirements Traceability Matrices (RTMs) failed because humans maintained links the system itself didn’t depend on
  • Provenance was overhead, not infrastructure — an afterthought with no enforcement mechanism
  • Intent-centric systems invert this: provenance becomes load-bearing
  • Agents cannot regenerate code without traversing the intent graph, so links are maintained automatically
  • The incentive structure shifts from “someone must remember to update the trace” to “the trace is how generation works”

Non-Determinism Challenges

  • Natural language specifications require canonicalization before serving as stable hash inputs
  • Agents surface implicit assumptions previously hidden in code
  • Non-deterministic generators may produce different outputs from identical intent graphs across runs
  • These are design constraints, not fundamental blockers: as specifications become more structured, ambiguity becomes visible rather than hidden in diffs
  • Evaluation layers (tests, contracts) absorb the variation that generators introduce — “probabilistic inside, deterministic at the edges”

Sources

  • Fowler, Chad (January 6, 2026). “Relocating Rigor.” The Phoenix Architecture (aicoding.leaflet.pub). https://aicoding.leaflet.pub/3mbrvhyye4k2e

    • Primary articulation of content-addressed intent graphs, specifications as executable inputs, and the unit of change being reasons not lines

  • Böckeler, Birgitta (October 15, 2025). “Understanding Spec-Driven-Development: Kiro, spec-kit, and Tessl.” Thoughtworks Technology Radar / martinfowler.com. https://martinfowler.com/articles/exploring-gen-ai/sdd-3-tools.html

    • Spec-as-source as an implementation approach; parallels to Model-Driven Development; non-determinism as key challenge

  • Swaminathan, Nikhil and Kartik Chivukula (December 3, 2025). “Introducing Powers.” Kiro Blog. https://kiro.dev/blog/introducing-powers/

    • Steering documents (POWER.md) as lightweight intent specification; dynamic context loading to prevent context rot

  • Ujcich, Benjamin E. et al. (2020). “Provenance for Intent-Based Networking.” IEEE International Conference on Network Softwarization (NetSoft). https://adambates.org/documents/Ujcich_Netsoft20.pdf

    • PROVINTENT framework: formal treatment of intent provenance in SDN/IBN; direct academic precursor to intent graph traceability

  • Mahmood, Yasir et al. (2023). “Intent-Driven Data Platforms: Replacing Pipeline Engineering with Declarative System Intent.” International Journal of Computational and Experimental Science and Engineering, Vol. 9. https://ijcesen.com/index.php/ijcesen/article/view/5092

    • Declarative execution model separating business intent from technical implementation; intent parser, execution planner, and runtime optimizer as architectural layers

Note

This content was drafted with assistance from AI tools for research, organization, and initial content generation. All final content has been reviewed, fact-checked, and edited by the author to ensure accuracy and alignment with the author’s intentions and perspective.

Graph View

Backlinks