Use Codebolt to write custom agents with your own loop, tools, models, memory, guardrails, and product logic.
Most agent products let you customize prompts. Codebolt lets you customize the agent itself. Write the loop in code, call the SDK directly, shape memory, route models per task, and package repeatable behavior as agents your team can run anywhere.
observe → plan → act
review → retry → commit
your logic at every step
Control planning, tool choice, memory reads, retries, reviews, commits, and handoffs. The agent is a program you own, not a hidden prompt behind a chat box.
agent.ts
→ codebolt.fs
→ codebolt.git
→ codebolt.memory
→ codebolt.llm
Use filesystem, git, terminal, browser, LLM, memory, task, notification, and orchestration modules from the Codebolt SDK. Agents can work inside the same engine as the CLI, Editor, Cloud, and server.
planner → GPT
coder → Claude
reviewer → Gemini
local tasks → Ollama
Use different providers for different parts of the work. Route planning, coding, review, extraction, and repetitive operations to the model that fits the job.
Start with a single specialist agent. Add reviewers, test agents, documentation agents, deployment agents, and monitors as the workflow grows. Because the agents run on the same engine, they can share memory, tasks, state, plugins, and environments.
custom-agent
↓
reviewer-agent
test-agent
docs-agent
deploy-agent
↓
coordinated by Codebolt