Immutable foundations

Values and pillars — the AIAD constitution

The philosophy of AIAD fits in one sentence: AIAD does not judge on effort, velocity or process compliance — it judges on the value realised for stakeholders. Everything else follows from it: four pillars, seven immutable values, three founding principles.

Judging on realised value

Without an explicit definition of "succeeding", teams optimise the wrong metrics. A technically perfect feature that nobody uses is a failure. A cycle completed at 100% with no measurable impact is a failure. A partial delivery that solves the real problem is a success. AIAD states this unambiguously so that the team never confuses activity with value.

The four pillars

1. Radical empiricism. Everything is a hypothesis until proven otherwise. Real data and usage take precedence over intuitions; pivoting quickly beats persevering in error; maximising learning speed matters more than production speed. Anti-pattern: "We know what our users want, we've been working with them for five years." Yesterday's certainties are tomorrow's failures.

2. Systemic orchestration. A team's value no longer lies in its ability to code, but in its ability to orchestrate AI agents. Humans define the why and the what; agents take charge of the how. Training one expert improves one person by 20%; properly configuring the AGENT-GUIDE improves the whole team by 50%. Anti-pattern: "Our best dev doesn't need agents." Their time would be better invested configuring the ecosystem for the other ten.

3. Fluidity through emergence. Artificial cadences — fixed-length cycles — create friction with the natural velocity of AI-assisted development. AIAD replaces them with a continuous flow adjusted to actual complexity: a bug is fixed in two hours, a feature takes three days, an exploration a week. Anti-pattern: "We keep our cycles but do AIAD inside them" — the worst of both worlds: rigidity plus confusion.

4. Intentional Excellence. Shipping code is not the goal; realising an intent worthy of the value invested is. The agent can contribute to technical excellence — it cannot carry the intent. The central question: "do the outcome criteria themselves reflect an intent worthy of the value invested?" The monthly Intent Workshop — a purely human space, no AI — is the mechanism that verifies this alignment. Anti-pattern: "We ship first, we'll improve later." "Later" never comes.

The seven immutable values

1. Primacy of Human Intent. The intent of what is built remains the exclusive domain of the human. The agent executes; it does not decide what deserves to be built.

2. Honesty about Contradictions. When constraints clash, AIAD tells the truth about the trade-offs rather than promising the impossible.

3. Intentional Sobriety. We build what is necessary, not what is impressive. Every feature must justify itself with a clear intent.

4. Radical Openness. AIAD is open source: methodological knowledge is shared, never locked away. Practices, tools and field feedback are published on aiad.ovh — transparency is a condition of collective improvement.

5. Uncompromising Empiricism. Hypotheses are validated through observation, not consensus. An AIAD team does not decide on the strength of an intuition when a measurement is possible.

6. Shared Responsibility. Success and failure belong to the team, not to an individual. Responsibilities are clear — blame never is. Field illustration (May 2026): an agent with financial autonomy (Cloudflare/Stripe protocol, $100 monthly cap) bought the domain superseal.cc instead of superseal.club — an error with no clearly designated owner upstream. Shared Responsibility requires that any delegation of autonomy to an agent, especially financial or operational autonomy, come with an explicitly assumed human responsibility and documented checkpoints (cap, audit log, rollback procedure). These checkpoints are detailed in the "Agents with irreversible effects" profile of the AIAD-AI-ACT agent and checked at the Execution Gate.

7. Human Authorship. Authorship of intent cannot be delegated. Every Intent Statement is written in the first person by an identifiable human; if an agent wrote it, the session is invalidated. External legitimation (2026): this value corresponds precisely to Martin Fowler's "serious agentic programming" — as opposed to vibe coding — and prevents cognitive debt (Fowler/Joshi, 2026). Field practice points the same way: across 29,585 pull request cycles analysed (arXiv, May 2026), merge authority remains "almost exclusively human", and frontier agents show a measured action bias of 35 to 65% that justifies a human checkpoint before execution. Since doctrine v1.9, the value unfolds into an authority continuum: each move from one authority level to the next is a distinct human act — and answering the agent's questions is not consenting.

The three founding principles of orchestration

Budget, delegation and dating (v1.9). Two 2026 developments change the calculation without changing the discipline. First, one-million-token context windows are becoming widespread: they push back the degradation threshold, they do not remove it. Second, some harnesses make subagents inherit the parent conversation: a delegation then costs the parent context on top of its own work, so it is better to delegate early. Cross-cutting rule: any capability or pricing assumption cited in an AIAD artifact is dated and revalidated periodically.

Positioning in the SDD ecosystem

AIAD is a meta-framework: the tools — Kiro (Amazon), GitHub Spec Kit, OpenSpec, Codeplain, SPDD and its REASONS Canvas, OpenAI Symphony, Claude Code Routines — handle the "how to work with agents"; AIAD handles the "how to work as a human-AI team": who carries which responsibility, how to validate that the intent remains intact, how to measure real value, how to maintain regulatory governance. It can integrate these tools as implementations of its spec layer.

The era of agent labs. In 2026, the major labs are repositioning themselves as "agent labs" and shipping their own harnesses (OpenAI Codex, Anthropic's Claude Code, Google's Gemini). This mass adoption creates a governance vacuum around human intent that harnesses do not fill: they optimise "how to generate", not "who wanted what, and who validates". AIAD is also model-agnostic by design: when a model is deprecated, Intent Statements, SPECs, the AGENT-GUIDE and the Execution Gate remain valid — only the executor changes. Governance lives in the artifacts, not in the vendor.

Industry validations 2026

The recent convergences are not follow-ups — they are independent validations of an earlier approach. Anthropic formalised Context Engineering as a discipline in its own right in March 2026 — AIAD has implemented it since v1.3. Thoughtworks Radar 2026 identifies Context Engineering, Harness Engineering and Spec-Driven Development as convergent concepts. R2Code (arXiv, April 2026) empirically validates spec-first: −41.7% of tokens and +7.4% fidelity. AWS Strands (2026) observes −96% of tokens on intent-based workflows. Gartner MQ 2026 projects 30-50% agentic productivity gains by 2028 — creating the governance vacuum that AIAD addresses structurally. Tang et al. (arXiv, June 2026) provide the founding empirical evidence: across 20,574 real coding-agent sessions, 91.49% require explicit human correction, and the seven forms of misalignment identified (misinterpretation of intent, scope creep, off-spec solutions, hallucinated requirements…) match the problems that the Intent Statement → SQS → Execution Gate cycle addresses structurally. At the end of July 2026, Cognizant added to its Flowsource product a module explicitly named "Spec-Driven Development": SDD is becoming an integrator offering for large enterprises — proof of market, not proof of effectiveness.

External support (v1.9). In July 2026, Anthropic published an overhaul of its context engineering doctrine: removal of about 80% of Claude Code's system prompt with no measured loss according to the vendor — a vendor figure, unaudited — and a recommended shift from simple instructions to rich, verifiable references (detailed test suites, existing code to port). This is an external validation of the SQS thesis: a good SPEC is not a longer list of constraints, it is a rich and verifiable artifact.