Kernel

To understand an agent runtime, follow one request all the way through. This series starts on the main path, then covers the pieces along it: messages and memory, the outer loop and compaction, model-stream adaptation, tools and permissions, instance boundaries, dual-channel streaming, plugins, and the SDK.

Articles in this series

Main path

• End-to-end request lifecycle: how user input becomes a tool call
• Message model and cross-turn memory
• runLoop and context compaction
• Processor and LLM adaptation

Capabilities and boundaries

• Tool system: registration, visibility, truncation, and permission gates
• Config and permission overlays (multi-layer merge)
• Instance boundaries and concurrency

External interfaces

• HTTP streaming and SSE dual channels
• How plugins work
• SDK-to-server flow

Writing conventions

This series is for readers who change the kernel or build deep integrations:

• State current Agent Forge mechanisms, boundaries, and verifiable checks; do not compare against third-party kernel product names in public docs
• Prefer reproducible minimal experiments or observable API behavior over slogans
• Deployment differences (for example SQLite vs Postgres) are product capabilities and configuration, stated directly