Code agent

Code agents need more than completions. They usually need repository access, iteration, memory, and some form of review or approval loop.

The Registry ships a code-reviewer agent you can install and point at your repo.

#Typical stack

npm install @agentskit/adapters @agentskit/runtime @agentskit/tools @agentskit/skills @agentskit/observability @agentskit/eval

#Recommended package mix

#What the architecture usually looks like

1. The agent receives a coding or review task.
2. Filesystem and shell tools inspect the repo state.
3. The runtime loops through investigation, edits, and validation.
4. Skills shape behavior for review, planning, or implementation tone.
5. Observability and evals help you keep the system trustworthy.

#Good defaults

• Separate read tools from write tools when you first ship.
• Add confirmation gates before destructive actions.
• Store structured traces so you can replay bad runs.
• Treat evals as part of the product, not an afterthought.

#Best follow-up guides

• Code assistant example
• Code reviewer recipe
• Mandatory sandbox
• Deterministic replay
• Prompt snapshots

#When AgentsKit is especially strong here

AgentsKit is a good fit when you want to combine tool use, structured runtime control, and post-run inspection without taking on a huge framework. The contracts also make it easier to split the system into safe layers as the agent becomes more capable.