intent.md committed by the person with the problemWhat Anthropic, Atlassian, Google Cloud and Shopify say about building software once agents write most of the code. Where they agree, where they genuinely disagree, and how much of it is actually measured.
Six stages, each ending in a committed file. Skills advise, hooks enforce.
Vendor playbook · Aug 2026Five workflow shifts standing on two foundations: a context graph and measurement.
Vendor playbook · Oct 2026A capability map across five phases, plus risks and best practices.
Marketing page · Sep 2026A five-year operating story: enablement layer, culture, and lessons learned the hard way.
Case study · 2026Four organisations with different incentives reach the same conclusions.
Code generation is no longer the constraint.The bottleneck moved to the work around the code: planning, review, maintenance.
Human review becomes the binding constraint.Anthropic, Atlassian and Shopify reach this independently. Google names it as a risk.
Context quality decides output quality.Not model choice. What the agent is told about your systems is the differentiator.
Humans stay accountable for judgement.Agents execute and enforce. People decide, approve and own.
It's an operating-model change, not a tool purchase.Explicit for Atlassian and Shopify, implied by Anthropic, not stated by Google.
When organisations competing for the same budget agree this precisely, the agreement is the finding.
intent.md committed by the person with the problemspec.md constrained by org skillsplan.md approved before any codeintent.mdNot differences in emphasis. Places where following one source means not following another.
Anthropic: markdown artifacts in git are authoritative; the commit history is the audit trail. Atlassian: one connected context layer across tools is authoritative, and the repo is one input.
An architecture decision, and expensive to reverse.
Shopify: "we don't know yet what's gonna win." Centralise the plumbing, not the choice.
Governance appetite decides this, not engineering preference.
Anthropic: hooks that block, settings engineers can't override. Shopify: iteration limits and spend alerts, deliberately no hard caps. Google: practices, no enforcement mechanism.
Regulated organisations will land near Anthropic. Know what it costs in developer friction.
Every number is a vendor claim. Directional, not a benchmark to commit to.
Scored on what you'd need to show an auditor. Three dots means explicit, implementable guidance.
My reading of the published material, not a vendor assessment.
Every stage commits a file the next stage reads; the commit history is the audit trail. Skills advise, hooks block. The part that survives contact with auditors.
Context is a platform investment, not a per-team effort. Measure before you scale, so you can tell real gains from hype.
Security flaws in generated code, IP exposure, comprehension debt, and the rule that generated code gets no exemption from review.
Nudge, showcase, make usage visible. Match the tool to the user. Protect comprehension as hard as you push adoption.