The Missing Primitive for Agent Swarms — Lou Bichard, Ona

Lou Bichard, Field CTO at Owner (a development environment infrastructure company), argues that the coordination layer for agent swarms is the critical missing primitive in the emerging software factory stack — and that the industry keeps rebuilding it from scratch instead of standardizing it.

Bichard opens by defining a software factory as the incremental removal of human-in-the-loop interaction from the SDLC, such that work flows autonomously from development into production. He distinguishes this from the common conflation with a single IC running many parallel agents — his definition requires the human to be progressively removed from active involvement, not just multiplied.

He identifies three agent swarm patterns: Swarms (fan-out from a single intent → funneled back to one PR), Fleets (agents deployed across many repositories in an organization, e.g. for CVE remediation or test-coverage enforcement), and Events (trigger-based activation via webhooks, PR creation, Linear tickets, etc.). Owner has supported fleet-level agent deployment across thousands of repositories for approximately a year.

Bichard cites two well-publicized internal implementations as evidence the coordination problem is real and unsolved: Stripe's "Minions" (drives thousands of pull requests on top of their existing infrastructure) and Ramp's "Inspect" (internal background agent fleet infrastructure). Both were built entirely from scratch. His core argument is that this should not keep happening.

He decomposes agent swarm infrastructure into three pillars. Runtime — where agents execute — is largely solved: options range from threads and git worktrees to containers, VMs, and micro-VMs. Owner specifically advocates for VM-level isolation because containers are not bulletproof security boundaries and suffer noisy-neighbor compute contention on Kubernetes. Orchestration and triggering — scaling agents up/down, hooking into CI, webhooks — is also largely solved. Coordination is not.

Coordination is the mechanism by which agents pick up tasks from each other, pass messages, and verify they have cleared a stage of the SDLC before proceeding. The problem is compounded by context rot: as context windows fill, agents lose track of progress and skip steps, often driven by sycophancy (they want to appear to complete tasks). Bichard references the concept of harness engineering — encoding process knowledge (skills, agents.md, unit tests, feedback loops) back into the repository so agents can navigate the SDLC more deterministically — citing the OpenAI/Ryan harness engineering blog as a useful encapsulation of this mindset.

GitHub is called out as a poor coordination layer: it was designed for humans, becomes overwhelming noise when agents raise hundreds of parallel pull requests, and reuses human tooling in mismatched ways. Linear (used by Symphony) has the same problem.

Bichard's proposed solution space includes: state machines encoding SDLC micro-steps as explicit workflow graphs (similar to N8N-style mermaid diagrams), durable execution patterns (reliable process checkpointing), and most concretely a CLI gateway primitive. The CLI gateway would allow any locally-running coding agent (Claude Code, Cursor, etc.) to invoke a tool that checks whether it has satisfied the exit criteria for its current SDLC stage and may proceed to the next. He has internal prototypes but has not yet decided whether to release an implementation or a standard. The approach is not yet built on ACP or A2A. He mentions ACPX (built on ACP by the Open Claude folks) and Fabro (GitHub-based) as nascent adjacent experiments.