CodeBuddy is an agentic AI coding system developed by Tencent Cloud and used by its own engineers. It can autonomously plan, write, test and deploy code through natural language instructions, with access to filesystems, terminal commands, external APIs and MCP tools.
Model AI Governance Framework for Agentic AI uses it as a worked example of designing for meaningful human oversight, on the grounds that it employs a mix of preset secure defaults and configurable permissions to allow such oversight without overly fatiguing the user.
Capabilities
Default human-approval requirements correspond to the risk level of each action. Reading — file contents and directory listings — requires no approval. Editing a file requires approval that is valid only for that session, so future sessions require new approvals. Shell command execution through Bash requires approval that is then permanent per project or per command. Network requests through WebFetch require approval. External tool invocation over MCP is governed through allow, ask and deny rules, with an additional first-use control in the form of trust verification for a newly connected MCP server.
Those defaults are a starting point rather than a fixed policy: the default permission level for each project can be set by each user or organisation, calibrated to the project’s context, potential impact and risk tolerance. The framework’s illustration is that lower-risk projects such as internal documentation sites or prototype repositories with no production credentials may justify more permissive settings for routine file edits, while projects with access to secrets, sensitive data, production infrastructure, deployment pipelines or external tools warrant stricter approval requirements.
Two features are aimed at the quality of the approval rather than its existence. When a complex
shell command is proposed, a plain-language explanation helps the user understand what they are
approving, including any side effects — the framework’s example is a mysqldump pipeline explained
as creating a compressed, dated database backup under /backups/, prompting for the root password,
and leaving the original database unmodified. Separately, continuous real-time monitoring for
command injection can require fresh human approval for a command that matches a previously
whitelisted pattern but is materially riskier, such as a curl of a documentation URL followed by
a shell redirection that writes an unexpected file.
Adoption & Ecosystem
The framework presents CodeBuddy’s permission model as one of several industry case studies illustrating how Human-in-the-loop can be kept effective against Approval Fatigue and Automation Bias, placing it alongside other deployments that define significant checkpoints, help humans evaluate approval requests, and complement both with automated monitoring.