The blog

Built in public, written the same way: notes on the protocol, the architecture, and what we're actually shipping.

Developers

The $400-Per-Bug Illusion: Benchmarking Long-Running AI Coding Agents on Real Multi-Hour Engineering Marathons

Benchmark long-running AI coding agents on 2,400-file production repos. See why CLI agents burn $400 per bug and how Imperal Cloud ends token waste.

long-running ai coding agents
Vall · 9 min
Protocol

Why Claude Code Loses Context After 40 Turns — And How Stateful Agent OS Architecture Ends Context Amnesia

Discover why Claude Code loses context after 40 turns due to terminal bloat and compaction, and how a stateful AI agent OS architecture prevents context amnesia.

AI Coding AgentsClaude CodeArchitecture
imperal.io Engineering · 8 min
imperal.io