Problem
Every team building LLM agents reimplements the same runtime primitives — retry classification, circuit breakers, usage and cost tracking, permissions, guardrails, checkpointing, and telemetry. The result is brittle, app-specific plumbing nobody else can reuse — and it quietly omits the auditability regulated teams now need.
Solution
techrevati-runtime provides those primitives as composable building blocks — usable standalone or through a single AgentSession. Version 0.4.0 adds an EU AI Act compliance kit (tamper-evident, hash-chained audit log; human-oversight controls; a residual-risk register; incident detection with deadline tracking; transparency reports), an MCP tool adapter, typed outputs, session memory with compaction, and step-level durability. It is a lightweight layer you embed into an existing Python stack — not a monolithic agent framework.
Outcome
Open IP that demonstrates our engineering standard — fully typed, zero runtime dependencies, Python 3.11–3.13, MIT licensed. Published on PyPI as techrevati-runtime and used internally across our agent systems. The compliance kit gives regulated teams implementation building blocks (not legal advice) mapped to specific EU AI Act articles. Still 0.x, so pin exact versions.