Loma

Most teams treat AI as a series of one-off conversations — a question answered, a draft written, a bug explained, all of it evaporating when the session ends. Loma is an open-source agent platform that intercepts that waste. It deploys agents across sales, support, product, and engineering functions, each one feeding a shared context layer — patterns, skills, tone signatures — so the next agent picking up a similar task starts with institutional memory rather than a blank prompt.

The context layer is Loma’s sharpest edge. When a developer fixes a webhook retry bug, Loma stores the diagnostic pattern. When support resolves a KMS timeout, that resolution becomes a reusable skill. The vendor describes this as 847 stored patterns, 23 skills, and 156 customers tracked in a live demo — numbers that illustrate the compounding mechanic even if they are illustrative. For RFP and infosec questionnaires specifically, the docs describe agents pulling from existing SOC 2 reports, past submissions, and ISO docs to answer the bulk of questions automatically, flagging only the ones requiring legal or executive sign-off.

Loma fits organizations that already run fragmented LLM subscriptions — some teams on Claude, others on GPT or Gemini — and want a single layer that routes tasks without locking into one provider. The Apache-2.0 license and self-hosted deployment mean your data and learned patterns stay on infrastructure you control, which matters when the context layer holds deal playbooks and customer health signals. The ceiling appears when you need enterprise-grade access controls, audit trails for compliance, or guaranteed uptime — the public documentation does not address these, and teams with procurement gates around SOC 2 or HIPAA will find precious little to hand to their security reviewers.

Slack is the primary control surface: agents proactively post briefs before meetings, accept task assignments via @mention, and surface flagged items for review inside channels your team already monitors. The vendor lists Docker and Kubernetes as supported deployment targets, and the model layer is described as swappable at any time — switching from a paid API model to a local open model is a configuration change, not a migration.