tagged
#ai
Retry is not a policy Building a coding agent Efferent retries only failures that can move, refuses multi-hour Retry-After sleeps, rejects empty 200s, and stops replaying a stream the moment content escapes. The model role, not the model id — and why 'code' earns its own tier Building a coding agent Smith routes refining, implementation, and one-shot helpers through three role-scoped views of one live settings store; the agent code never learns a provider name. Subscription OAuth is the hard half of multi-provider auth An Anthropic subscription token changes the HTTP request, the first system block, refresh persistence, and the login race — it is not an API key with a different prefix. Pay for the index, not the book: extending an agent with markdown on disk Building a coding agent Smith exposes skill names up front, loads bodies on demand, lets workspaces shadow bundled procedures, and graduates corroborated memory into the same mechanism. Error messages are prompt engineering now Building a coding agent In Efferent, tool failures correct the current turn and gate findings brief the next attempt. The placement, shape, and recovery verb matter more than the exception text. The context window is a resource you engineer, not a buffer you fill How an agent remembers Efferent folds context at two different boundaries, preserves a safe tail, and deterministically reattaches the rules a lossy summary must never be trusted to remember. Prompt caching is a design property, not a billing detail An agent re-sends 95% of its prompt every turn. Whether you pay full price again is designed, not billed. The transcript is the memory How an agent remembers The model remembers nothing between calls. The useful engineering is in the log: atomic positions, decoded reads, non-destructive folds, and transactional branches. An agent loop is a while loop with good manners Building a coding agent One agent turn, dissected: exits, breakers, and recovery paths are the actual engineering — in a loop whose own codebase bans the while statement. AI applications are Effect-shaped Effect, from zero Flaky IO, string boundaries, provider churn, fan-out, cancellation that stops billing — the defining problems of LLM apps are the ones effect systems were built for.