what i'm on, right now.
shipping
finished the agentic security quartet: hider (tiered redaction gateway, two live tiers, 130 tests, fail-closed in flight), llm-intelligence-layer (advisory vs acting agent modes with human-approval holds as graph interrupts), arbiter (finding fusion ranked by blast radius behind a deterministic guard), and a guard layer for aissist with provider-tier data classification and a pr template that replaced per-team security review. applying for lead agentic systems seats with all four public.
logs provider
| redact in flight ^
v |
+--------------------------------------+
| tier1 patterns -> tier2 ner -> gate |
| degrade? refuse. nothing crosses. |
+--------------------------------------+ writing
the llm-batcher series (translation, microbatching, self-observation) is done and live. next up: a write-up on why "100% of mutating tool calls are gated" was true only by accident in my own agent runtime, and what enforcing it in the executor taught me about agent safety.
openai client claude
| POST /v1/chat/completions ^
v | /v1/messages
+---------------------------------------------------+
| llm-batcher |
| translate | microbatch | observe |
+---------------------------------------------------+
|
GET /metrics -> p50 / p95 / p99 + $cost