Code in the terminal with an agent

The poster uses Grok as a terminal coding agent that reads, edits, runs tools and spawns sub-agents.

What They Did

The poster says they built agenc-core as a terminal agent harness where Grok drives the work directly in a CLI/TUI: it reads files, edits code, runs tools, and can spawn parallel sub-agents with separate context, tools, and token budgets. In the example they watched it split a divisible job into eight workers to build eight separate C terminal games at once, while the main orchestrator integrated and verified the results and a human handled approval prompts for tool calls.

How It Went

The post presents it as working well enough to be the author’s showcase, including the eight-game parallel build and integrations like Ledger actions gated by on-device approval. It does not give benchmark numbers, failure rates, or a detailed comparison against other coding agents.

Worth Knowing

The main catch is the permission model: tool calls can require human approval, and any Ledger action that moves value needs physical confirmation on the device.