AI agent governance: four platforms, four philosophies
AI agent governance is the set of tools, policies, and controls that keep non‑deterministic AI agents predictable enough for production, covering identity, security, audit, testing, and change management so enterprises can trust agents with real users and data at scale. Today, four platforms stand out with distinct views on enterprise AI control: Raft turns agents into team members in shared workspaces, Loom focuses on identity‑driven security on AWS, Harness wraps agents in deterministic delivery pipelines, and Presence concentrates on policy‑rich customer and employee support. In plain terms, Raft is ideal for agent‑native builders, Loom for AWS platform teams, Harness for software delivery leaders, and Presence for enterprises that want voice and chat agents tied to strict guardrails. The right choice depends less on model quality and more on how much control, visibility, and compliance you need over production AI agents.
| Spec | Raft 1.0 | AWS Loom / Harness DLC / OpenAI Presence |
|---|---|---|
| Primary focus | Multi‑agent collaboration with persistent memory and parallel task execution in shared human‑agent workspaces | Loom: identity and security for AI agents on AWS with RFC 8693 token exchange; Harness DLC: deterministic pipelines, evals, and security for production AI agents; Presence: voice and chat agents with built‑in policies, guardrails, and guardrail updates |
| Deployment model | Agents run as persistent processes on the user’s own hardware via a lightweight daemon, keeping compute, code, and data under user control | Loom: pre‑written, configurable Python agents on Amazon Bedrock AgentCore Runtime with configuration‑only changes; Harness DLC: integrated into existing SDLC pipelines (details not fully specified); Presence: limited general availability program with supported deployments, not self‑service |
| Identity & security approach | Local execution minimizes data exposure; identity is per‑agent, but enterprise identity propagation is not the core feature described | Loom: implements full authorization code flow and RFC 8693 token exchange so user and agent identity travel through delegated chains; Harness DLC: adds governance, testing, and security used for application code to agents; Presence: enforces policies, guardrails, approved actions, and guardrail interventions for enterprise data protection |
| Determinism & evals | Relies on agents reviewing each other’s work in shared threads to catch issues before humans see them | Loom: avoids runtime code generation, keeping agent code constant between deployments; Harness DLC: uses quality gates and eval scores on correctness, safety, and performance wired directly into pipelines; Presence: simulations and graders test agents against common, edge, and high‑risk scenarios before launch |
| Pricing / availability | Free tier with channels, tasks, local agents, and 30 days of history, plus a Pro plan at USD 8.80 (approx. RM41) per seat per month billed annually, and an enterprise tier listed as coming soon | Loom: open‑source reference platform on AWS Labs, not a managed service; Harness DLC: released as part of broader AI agent governance capabilities (pricing not specified); Presence: limited general availability via enterprise account teams, not self‑service and pricing not specified |
Raft 1.0: shared workspaces for agent‑native teams
Raft 1.0 treats AI agents as full team members, giving them persistent identities, memory, and expertise inside shared workspaces of channels, threads, tasks, and mentions where humans and agents work side by side. Unlike ad‑hoc scripts or chat tabs, each agent runs as a persistent process on whichever runtime fits—Claude, Codex, Hermes—and executes on the user’s own hardware via a lightweight daemon, keeping compute, code, and data under the user’s control. Agents claim tasks, run in parallel, hand work to one another, and review each other’s output in shared threads so independent agents cover independent blind spots before a human steps in. Raft’s free tier includes channels, tasks, agents on local machines, and 30 days of message history, while a Pro plan costs USD 8.80 (approx. RM41) per seat per month billed annually and an enterprise tier with private deployment and SSO is listed as coming soon. This makes Raft compelling for engineering‑heavy teams that want agent collaboration and data control more than formal enterprise AI governance.
AWS Loom: identity‑first enterprise AI control on AWS
Loom is an open‑source, opinionated agent platform that shows platform teams how to build, deploy, and govern AI agents on AWS with security controls baked in from the start. It builds agents with the Strands Agents SDK, runs them on Amazon Bedrock AgentCore Runtime, and packages a management UI and backend API with identity provider integration, scope‑based authorization, and lifecycle management for agents, memory resources, MCP servers, and agent‑to‑agent integrations. Its standout feature is identity propagation across delegated actor chains: when an agent calls an MCP server that then calls a REST API, Loom implements the full authorization code flow and RFC 8693 token exchange so both user and agent identity travel in downstream access tokens while the delegation chain stays intact. This keeps downstream systems limited to data the originating user can access. The deployment model deliberately avoids runtime code generation: teams deploy a pre‑written, configurable Python agent and inject guidelines, memory, and configurations at deployment time, so code stays constant while configuration changes. Loom is best for cloud platform engineers who need strict identity and security for production AI agents without handing over control to a managed service.
Harness AI Agent DLC: deterministic pipelines for non‑deterministic agents
Harness’s AI Agent Development Lifecycle (DLC) starts from a blunt truth: agents change their answers, making incidents hard to reproduce and traditional testing unreliable. Instead of forcing agents to become deterministic, Harness brings deterministic mechanisms and governance to the pipeline that surrounds them. The new service lets developers ship AI agents with the same governance, testing, and security they already rely on for application code. As Trevor Stuart explains, “Don’t think about trying to make the agent predictable; instead, make the pipeline around it predictable… grade the response on correctness, safety, and performance, and wire that score straight into the pipeline as a pass‑fail gate”. Harness AI Evals makes agent quality measurable by letting teams define eval datasets, hook up scoring functions, and set quality gates that automatically catch regressions whenever an agent or model changes. For organizations with mature CI/CD, DLC is the natural choice to extend existing delivery controls to production AI agents without pretending their outputs are reproducible.
OpenAI Presence: policy‑rich voice and chat agents with guardrails
Presence is OpenAI’s deployment platform for AI agents that handle customer support and internal service requests across voice and chat, with enterprise AI control at its core. Rather than being a standalone model, Presence brings together policies and standard operating procedures, guardrails, approved actions, simulations, evaluation tools, and a Codex‑powered improvement process into one production stack. Before an agent goes live, teams can test it against common requests, edge cases, and higher‑risk scenarios; simulations and graders check whether it reaches the correct outcome, follows policy, uses tools properly, and escalates when needed. Guardrails intervene whenever an interaction moves outside the company’s boundaries, protecting enterprise data and compliance. After launch, Codex reviews real sessions and escalations and suggests improvements that staff test and approve before they go live, allowing guardrails and behaviors to evolve without rewriting the agent from scratch. Presence already runs OpenAI’s own English‑language phone support and resolves 75% of inbound calls without human help. It suits enterprises prioritizing policy enforcement, voice and chat experiences, and ongoing guardrail updates over direct, hands‑on control of the underlying models.
Buy if / Skip if
- Buy the Raft 1.0 platform if you are an agent‑native engineering team that wants persistent, collaborative agents running on your own hardware with a cost‑effective Pro tier and generous free access.
- Skip the Raft 1.0 platform if your top priority is formal enterprise AI governance, identity propagation across systems, and compliance frameworks rather than multi‑agent collaboration.
- Buy the AWS Loom platform if you are an AWS platform or security team that needs deep identity‑driven AI agent governance, RFC 8693 token exchange, and configuration‑only deployments on Bedrock AgentCore Runtime.
- Skip the AWS Loom platform if you want a fully managed agent service instead of an open‑source reference implementation that your team must operate and extend.
- Buy the Harness AI Agent DLC platform if you already depend on CI/CD and want deterministic pipelines, eval‑driven quality gates, and familiar governance and security controls around unpredictable production AI agents.
- Skip the Harness AI Agent DLC platform if your organization lacks established delivery pipelines or is not ready to treat agents like first‑class software artifacts in its SDLC.
- Buy the OpenAI Presence platform if you need voice and chat agents for support use cases with built‑in policies, guardrails, simulations, and Codex‑driven guardrail updates protecting enterprise data.
- Skip the OpenAI Presence platform if you require self‑service deployment or want to host and tune your own agents without going through an enterprise account team.
