A Non-Exhaustive Inventory of exe's Software Factory

exe.dev has assembled a suite of agents that each address a specific operational need, from security scanning to rollout monitoring. The security agent runs automatically and prioritizes recent modifications, cutting down on manual review effort. To adopt this, teams must integrate the agent with their version-control system and ensure it can access change metadata. The approach may be less effective for projects that lack clear change timestamps or that enforce different security review processes.

The alert-investigation agent, Sisyphus, consumes alerts from paging systems and Slack, then cross-references logs, metrics, and source code to provide context for engineers. This reduces the time spent digging through raw data when an alert fires. Implementing Sisyphus-like functionality requires routing all relevant observability streams into the agent and configuring the notification channels it should monitor. If logs are unstructured or the alert volume overwhelms the agent, its usefulness could degrade.

Athena supervises infrastructure deployments by watching runtime metrics and logs, alerting operators if a rollout deviates from expected behavior. By automating this watch, operators no longer need to manually track long-running deployments. Deploying Athena demands hooks into the deployment pipeline and access to the same metric collection used by the rest of the system. In environments where deployments are instantaneous or where metric endpoints are unavailable, Athena would have little to monitor.

Additional bots handle flaky or slow CI tests, self-healing UI tests via language models, and daily summaries of git activity and help-thread activity. These agents automate routine CI maintenance and keep stakeholders informed without manual effort. Adoption costs include granting the bots read access to CI logs, test artifacts, and, for the UI tests, access to an LLM service. Their effectiveness drops if test logs are incomplete, the LLM API is unreachable, or the summary channels are misconfigured.

All of the agents run on isolated VMs at exe.dev, deliberately restricting the tools they can invoke to mitigate the "Lethal Trifecta" of private data, untrusted content, and external communication. Engineers replicating this model must provision similar isolation and enforce strict capability boundaries for any bot that processes external inputs. This strategy may not be viable in environments where VM isolation is impractical or where the required tooling cannot be sandboxed.