How we built groundcover Investigations – getting agent swarms to follow the scientific method
Blog post from Groundcover
groundcover Investigations organizes incident response as a shared, evidence-based tree of questions, hypotheses, findings, and assessments that can be updated by engineers, groundcover agents, and external coding agents. To coordinate these participants, groundcover moved investigation storage and rule enforcement into a dedicated service, separate from the AI agents, and assigned invariant rules—such as preventing duplicate investigations and controlling status changes—to code, while agents handle judgment-intensive tasks like forming hypotheses and evaluating evidence. Structured API schemas, restricted tools, and a moderator agent help ensure that proposals are valid, traceable, and supported by evidence, while concise updates guide agents through changes made by others. Monitors, Slack, automations, and MCP-connected agents all resolve to the same investigation record, with brief assessments posted to Slack and full details kept in groundcover. The team’s broader lessons are to define the data model before tuning prompts and enforce essential rules in code; future work will explore a bottom-up approach in which agents gather evidence first and identify potential hypotheses from it.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| LLM | 7 | 2,788 | 618 | 179 | -58% |
| MCP | 7 | 7,404 | 711 | 183 | -38% |
| Agent Evaluations | 4 | 92 | 42 | 25 | -45% |
| AI Agents | 3 | 4,689 | 923 | 218 | -33% |
| Harness engineering | 2 | 87 | 52 | 37 | -43% |
| Jev | 1 | 349 | 31 | 16 | -73% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.