The Savra Blog

Jev from TypeSafe AI answers yes/no, pick-one and rate-on-a-scale questions in 70 to 500ms for $0.042 per million input tokens. What it is, how the API works, and where it fits in an agent stack that already went single-agent.

A 136-entry bug ledger with no ticket system: what every entry carries, the eight buckets a census found, and the four-verdict triage that keeps it honest.

A production supervisor-and-specialists build: routing decisions, per-agent tool registries, 8,000 tokens down to 800, fallback models for rate limits, and streaming the agent's reasoning to the user.

No review, self-critique, supervisor review, human approval and double review, with the latency and token cost of each, and a decision matrix that picks by stakes and speed.

A zero-LLM heuristic classifier that picks an execution pattern in under 5ms, why asyncio.gather beats sub-agents until it does not, and when to expose pattern control to users.

Three execution patterns, chosen per query rather than per system: unified sequential for 80%, unified with parallel tools for 15%, multi-agent parallel for 5%. Plus why skills are orthogonal to all three.

A routing LLM call added 300ms and doubled token overhead on every request across 10,000 production queries. Why skills beat specialists for sequential work, and what removing the router changed.

Define agents as data so the directory and API stay in step, then trace every run. What per-trace token and cost attribution makes visible, and why it comes before optimisation.

A zero-tool supervisor routing to specialised agents cut tokens per request from 8,000 to 800 and took throughput from 2 requests a minute to 12. What the pattern is, and what we later learned it costs.

Multi-agent burns about 15x the tokens by design and carries 14 catalogued failure modes. Why one agent with dynamic skills, curated context and tiered memory is the 2026 default, and the read-write test for the exception.

Pydantic AI tool calls inside a LangGraph node do not emit to astream_events. Capture state from the node's return value via on_chain_end instead, in about ten lines.

A field report on a single deep-research run: where 106 agents and $76 went, why 61% of the bill was cache writes, and the one thing adversarial verification bought that a $1 query cannot.