Writing about AI engineering, agents, and reliable software systems.
Anthropic found a small, privileged workspace inside Claude — internal patterns, readable with a new Jacobian 'J-lens', holding concepts the model reasons with but never writes down. Not a consciousness claim. Maybe something more useful: a window into the thoughts a model doesn't print.
Sol, Terra and Luna offer different capability and cost tradeoffs. A current API comparison of reasoning controls, token prices and workload evaluation.
Anthropic's ClaudeDevs shared the pattern they lean on with Fable 5: don't make the frontier model do the work — make it the advisor. A cheaper executor runs the loop and calls up to Fable 5 only at the decisions that matter. Most tokens bill at the cheap rate; judgment stays near the top.
GPT-Live combines full-duplex voice with backend model delegation. How live conversation, tool permissions, interruptions and application state fit together.
Anthropic’s oral history traces Claude Code from early coding assistants and the internal clide CLI to a public product. Lessons in model capability, product design and feedback.
An illustrative PATH wrapper selects an authorized Claude Code identity and model at process launch. Its scope, inheritance limits and failure policy explained.
How an autonomous coding loop can discover missing work, track dependencies and verify live behavior. A hypothetical architecture with explicit authority, budgets and completion checks.
Anthropic's most capable widely released model was frozen by a US export-control order three days after launch. Eighteen days later the controls came off — and the reason it came back says more than the reason it left. The whole story, in one place.
Reliable handoffs combine explicit design requirements, component mappings and rendered verification. Executable contracts can reduce repetitive checking for covered requirements; review still depends on model capability, integration quality and untested behavior.
An architecture for evaluating autonomous coding loops: durable state, bounded workers, independent acceptance evidence and honest human handoffs.
There are two ways to use a coding agent: sit in its terminal, or take its face off and drive the engine from your own code. If you're building anything around agents, the second mode is the whole game — and it's full of traps nobody warns you about.
Ralph-style loops launch fresh workers and keep state in files. How they differ from an in-session Stop-hook plugin, and what verification and isolation they require.