Architecture Intelligence for the AI-Development Era

Abstract

AI coding agents now generate the majority of code on fast-moving teams — faster than any human architect can review, document, or govern. The discipline that kept systems intentional (upfront design, hand-written ADRs, periodic diagrams) assumes a human authoring pace that no longer exists.

ArchSteer proposes a different model: the system authors architecture artifacts continuously from code, and the architect curates them. This inverts authorship and makes architecture continuously observable, governable, and auditable.

1. The shift: docs are agent infrastructure

Persistent design context — a current architecture map plus decision rationale — is the primary lever against agent context degradation, the mechanism behind structural “slop.” An agent scanning a half-migrated repo will replicate the dominant local pattern, because that is the strongest signal in its context.

Documentation has therefore stopped being human paperwork and become critical infrastructure that AI agents consume. If it is stale, agents are steered wrong by default.

2. The core principle: invert authorship

Authored artifacts are always behind reality; emitted artifacts are reality. ArchSteer derives a single typed model from source and treats every downstream artifact — diagrams, ADRs, conformance reports, the evolution feed — as a projection of it.

The architect’s role changes from author to editor-in-chief: they ratify decisions the system surfaces and declare the invariants they care about, rather than maintaining documents by hand.

3. Two planes: descriptive and prescriptive

To be a must-have for every enterprise — not only those mid-migration — ArchSteer separates two planes. The reality plane is descriptive and needs zero setup: it answers “what is our architecture, and how did it change?” This is universal, because nobody knows their real architecture anymore.

The intent plane is prescriptive and opt-in: the architect declares targets and ArchSteer enforces them and steers agents toward them. Observability is the vitamin everyone takes; governance is the painkiller they pay for.

4. The four pillars of Architecture Intelligence

5. How agent steering actually works

Steering is a delivery mechanism, not a slogan. ArchSteer compiles the relevant slice of the architecture model plus the governing ADRs into file-scoped guardrails and writes them directly into the agent’s context files — CLAUDE.md, AGENTS.md, and .cursor/rules. Because the directives are scoped to the files in play and reference the target ADR instead of inlining the whole model, they survive the context window.

For agents that speak the Model Context Protocol (MCP), an ArchSteer MCP server exposes the live model and intent on demand — get_target_pattern(file), check_file(file), current_architecture() — so an agent can query the target pattern mid-edit rather than guessing from adjacent legacy code. The same model that steers is the model that CI checks, so what you steer toward is exactly what gets enforced.

6. Why this is a category, not a feature

Adjacent tools each own one slice — and each was built for a human authoring pace. Traditional static analysis and single-language linters throw errors but don’t steer AI agents. Manual C4 and diagram tools are dead on arrival when agents ship hundreds of lines a minute. Migration-recipe engines automate refactors but don’t govern what new code does next. AI code reviewers comment after the fact, on one diff at a time.

None fuse a live code-derived model with auto-built ADRs, a net-new conformance ratchet, an evolution timeline, and AI-agent steering — local-first and language-agnostic. That fusion is Architecture Intelligence: observability + governance + decision provenance for the AI era.

7. Go-to-market: the Drift Index bridges free to paid

Land with a free, read-only X-ray: run it on one repo, see current state plus recent evolution plus drafted ADRs. Expand to governance priced per repo and per architect seat: CI gating, agent steering, multi-repo rollups, and the executive Drift Index dashboard.

The Drift Index is the bridge. The moment a team watches their drift score on the free tier, the question changes from “what is our architecture?” to “how do we stop the bleeding?” — and stopping it requires the governance tier. Lock-in follows naturally: ArchSteer becomes both the source of truth and the decision ledger, and removing it means going blind and losing the audit trail.

8. The must-have test

The bar a principal architect applies: would the Chief Architect feel exposed at a board or audit review without it? Slides and wikis fail that test. An always-current, code-derived system of architectural state and decisions passes it.