https://arble.org/blog.html Engineering, research and product writing from Arble. en https://arble.org/blog.html#latest https://arble.org/blog.html#latest A context window is working memory, not storage. What retrieval looks like when the unit is a fact with provenance rather than a chunk of text. https://arble.org/blog.html#latest https://arble.org/blog.html#latest Registry lookup was 400 ms at the ninety-ninth percentile once workspaces passed a few hundred tools. What we changed, and why the obvious cache was the wrong fix. https://arble.org/blog.html#latest https://arble.org/blog.html#latest Every write stops to ask. The hard part is not the prompt — it is deciding what counts as one decision, and how to avoid training people to approve without reading. https://arble.org/blog.html#latest https://arble.org/blog.html#latest From JSON schema to a validated call and back into the transcript, including the failure modes that only appear once a model can call forty things. https://arble.org/blog.html#latest https://arble.org/blog.html#latest What a session owns, what it borrows, and how compaction preserves the instructions that set it up rather than the small talk that followed. https://arble.org/blog.html#latest https://arble.org/blog.html#latest Decorators, generated schemas, streaming results and publishing, in four languages. Why the SDK generates the schema instead of asking you to write it. https://arble.org/blog.html#latest https://arble.org/blog.html#latest Multi-agent work looks like a distributed system because it is one. What we borrowed from actor models, and where the analogy breaks. https://arble.org/blog.html#latest https://arble.org/blog.html#latest Latency is the smaller argument. The larger one is that an agent holding your keys and your context is a different trust proposition to one holding a copy. https://arble.org/blog.html#latest https://arble.org/blog.html#latest When to write a tool into the runtime and when to reach for a protocol server, measured on latency, typing, failure handling and who you are trusting. https://arble.org/blog.html#latest https://arble.org/blog.html#latest A trace that stops at the HTTP boundary tells you nothing about why an agent chose a tool. Instrumenting the loop itself. https://arble.org/blog.html#latest https://arble.org/blog.html#latest Pairing, key exchange and the relay that carries ciphertext it cannot read. Plus the host-key problem we have not solved yet. https://arble.org/blog.html#latest https://arble.org/blog.html#latest Two words used interchangeably that describe different mechanisms with different failure modes. Why conflating them produces agents that think beautifully and do nothing. https://arble.org/blog.html#latest https://arble.org/blog.html#latest Readiness against liveness, draining a node mid-run, and why additive migrations are what make a rolling upgrade possible at all. https://arble.org/blog.html#latest https://arble.org/blog.html#latest Designing a command surface that is pleasant to type and safe to script — including why every command has a --json mode. https://arble.org/blog.html#latest https://arble.org/blog.html#latest Token boundaries, backpressure, and the disconnect bug that silently discarded work for two releases before we understood it.