How We Built a React Native News Agent With Eve
Featuring
Turning Repository Activity Into Useful Updates
During the stream, Kewin Wereszczyński and Hur Ali built an Eve agent that collects recent activity from selected GitHub repositories and turns it into a focused daily digest. Instead of forwarding every update, the agent filters out routine CI, test, style, and maintenance commits, highlighting changes related to features, fixes, refactors, and performance.
The implementation focused on the React and React Native repositories, but the same structure can be extended to Expo, Metro, Hermes, or repositories maintained by an individual team.
Combining Deterministic Tools With Agent Reasoning
The agent used an OpenAI-compatible custom model made by Callstack (called Apex) and a small set of clearly defined Eve tools. GitHub data was retrieved through Octokit, with shared repository logic extracted into a reusable function.
Separate tools were exposed for React and React Native. This kept tool selection explicit and allowed the agent to request data only from the repositories relevant to a prompt. A small tool factory reduced duplicated code without giving the model unnecessary control over repository selection.
The stream also showed where deterministic code and model reasoning fit best. Code handled data retrieval and integration boundaries, while the model categorized commits, removed low-value updates, generated short explanations, and formatted the final digest.
Encoding the Output as an Eve Skill
Rather than writing a large set of transformation rules for every possible commit format, the agent used an Eve skill to describe how repository activity should be processed.
The skill instructed the agent to group changes by repository and area of impact, generate a short description for each relevant commit, preserve links to the source, and report how many maintenance-focused commits were omitted.
This made the output easier to adapt than a rigid parser while keeping the agent’s responsibilities narrow. The approach works well for a low-risk information workflow, but the speakers also stressed that an agent framework does not remove the need to account for hallucinations, context loss, and model behavior.
Deploying the Agent and Publishing to Slack
After testing the agent through Eve’s CLI and web channel, the project was deployed to Vercel. A scheduled Eve task was configured to prompt the deployed agent for the latest repository activity, allowing it to run without a developer starting the project manually.
The agent was then connected to a Slack workspace through a Vercel connector. Each digest begins with a main Slack message, while updates for individual repositories are published as separate replies in its thread.
The team manually triggered the deployed schedule during the stream and verified the complete path: Eve started the run, fetched repository activity, processed the commits, and published the structured digest to Slack.
Choosing the Right Work for an Agent
Eve becomes useful when the workflow benefits from interpretation: removing changing forms of noise, explaining unfamiliar commits, adjusting the output to different repositories, or adding new tools and communication channels over time.
The practical takeaway is to keep deterministic operations in code and use the model where judgment adds value. Eve provides the structure for combining those two parts, along with channels, schedules, sandboxed execution, and visibility into agent runs.
Watch the full live stream and follow the implementation step by step.
