The context-budgeting post earlier in this stretch made the qualitative case. This closes the loop with actual measured cost curves, since "long context is more expensive" undersells how non-linear that cost actually is."
Most schema validation discussion focuses on validating what the model produces. Validating what flows INTO the agent — tool results, retrieved content — at the boundary matters just as much and gets far less attention.
Streaming works naturally for prose — show tokens as they arrive. A structured JSON output streamed the same naive way produces invalid, unparseable partial JSON at every point except the very end.
Retrying a failed tool call is standard reliability practice for read operations and dangerous for write operations, unless the tool call is genuinely idempotent — which most agent tools aren't, by default.
A malformed function call from the model isn't a rare edge case — it's a routine occurrence at scale, and how the system responds determines whether it's a minor blip or a cascading failure.
The versioning discipline covered for skills and MCP tools earlier in this blog applies to output schemas too, with a wrinkle specific to structured outputs — the model itself needs to reliably produce the right shape, not just the schema definition being correct.
"Just add more context" isn't a memory system. Here's what production agent memory actually requires — explicit tiers, controlled write policies, and retrieval that scales without constant human curation.
Trust and identity, covered in the previous post, are the entry condition for a cross-org handoff. What happens during and after the handoff has its own set of practical failure modes worth naming explicitly.
Delegating a task to an agent your own team built is a different trust problem than delegating to one owned by another company entirely. A2A interoperability needs an actual answer for the second case.
An A2A agent card is the interface another agent uses to decide whether and how to delegate to yours. An inaccurate one causes failures that only show up after a delegation attempt, not before.