Kong has written up how it runs coding agents inside its own engineering organization, and the design choices are the useful part. In How Kong Built an Agentic Engineering Platform, the company describes KSAI, Kong Skills for Agentic Intelligence, which “brings AI-assisted code review and implementation into GitHub, where our engineering work is actually done.” The reason it exists is familiar: “Previously everyone was building something for their own, using their skills/prompts and tricks.” The agents run in CI, “so closing a laptop doesn’t stop the work,” and all model traffic goes through Kong’s own AI gateway for “observability governance” and to “hide LLM providers behind a single abstraction.”
The one figure comes from an A/B test the post describes carefully: on 10 percent of runs, Claude Code and OpenCode reviewed the same change with the same model and effort, and “we didn’t see much of a difference in the cost of Opus runs, but surprisingly Sonnet runs were 40% cheaper on OpenCode.” The isolation is the part to copy. OpenCode runs in a bubblewrap sandbox with a read-only root filesystem, an isolated temp directory and separate namespaces; GitHub tokens, OIDC variables and runner credentials are “removed from the agent’s environment”; the model token is brokered outside the sandbox through a read-only file. “Authorisation and approval remain separate boundaries”: a control plane checks who may start a paid or writing run, and the pull request author and co-authors cannot approve their own changes. The candid lesson concerns the first design, where each subagent explored the codebase separately: “In effect we were burning tokens multiple times for the same thing.”
The catalog holds the products this dogfoods. The Kong provider page lists 139 API pages, and KSAI’s model traffic runs through the Kong AI Gateway, whose per-request cost attribution this series covered last week, with agent traffic governed by the Kong Agent Gateway. The agentic access profile maps 1,242 operations, 782 of them acting and 122 human-in-the-loop.
The Kin Score is 71.8, exemplar band, and Agent Readiness is 46.9, agent-ready. The MCP server dimension is still unlit, the regression filed on our side earlier this month. The post’s security design is a list of the dimensions this series measures, applied internally: separate identity from approval, keep credentials out of the agent’s reach, and attribute every model call. Delegated identity is lit on Kong’s record, consent identity and dry-run mode are not. Kong has built for its own engineers the controls it sells to customers, and said which first attempt wasted tokens. That is the kind of write-up other platform teams can actually reuse.