What changed in late 2025

• Microsoft Agent 365 brings centralized oversight (authorize, quarantine, secure agents; third‑party coverage) and lands in early access via Ignite. Reuters.
• OpenAI AgentKit focuses on fast prototyping to production with a connector registry and agent evals; OpenAI also added MCP (Model Context Protocol) support in core APIs and Realtime for voice agents. TechCrunch, OpenAI.
• Salesforce Agentforce 360 deepens enterprise agent workflows, Slack integration, and a Builder for deploy‑test cycles. TechCrunch.
• Google Vertex AI Agent Builder/Engine adds A2A (Agent‑to‑Agent) interoperability, code execution sandbox, memory bank, and managed runtime (GA/updates in 2025). Google Cloud, Release notes.
• Amazon Nova Act (research preview) is a browser‑control agent + SDK for developers. TechCrunch.
• Notion Agents bring agentic workflows into a popular productivity stack for data analysis/task automation. TechCrunch.

Related reads on HireNinja for deeper implementation details:

• Microsoft Agent 365: what it means and a 14‑day prep plan
• Stop Agent Sprawl: ship an agent registry + RBAC in 7 days
• Agent observability blueprint (traces, evals, audit)
• Connect AgentKit + Agentforce 360 + MCP/A2A in 14 days

The evaluation rubric (use this for your RFP)

1. Identity & RBAC — first‑class agent identity, credential isolation, OAuth/OIDC patterns.
2. Registry & approvals — catalog of approved agents/tools, change control, and audit trails.
3. Interoperability — support for MCP and A2A to avoid lock‑in and enable multi‑agent handoffs. OpenAI MCP, Google A2A.
4. Tooling & connectors — native connectors, search/RAG, code execution sandbox, SIP/voice if you need phone agents. OpenAI tools & Realtime, Vertex updates.
5. Observability — traces (OpenTelemetry‑style), evals, step logs, red‑flag alerts.
6. Safety & compliance — PII handling, policy checks, PCI/PSD2 scope for checkout, EU AI Act readiness.
7. Performance — latency, success rate on tool use, fallbacks, cost per successful task.
8. Workflow fit — channels (web, chat, email, voice), human‑in‑the‑loop, and escalation.
9. Data & residency — regional control and model routing options.
10. TCO — platform fees + model costs + logging/observability + security review.

For governance controls to pair with this rubric, start with our 2025 Agent Governance Checklist.

Platform snapshots (fast facts for shortlisting)

OpenAI AgentKit

• Best for: speed from prototype to production; strong evals; growing connector registry; Realtime voice agents.
• Interoperability: supports MCP (remote servers) across Responses and Realtime APIs. OpenAI.
• Consider: governance add‑ons (registry/RBAC/observability) and cost controls at scale. TechCrunch.

Salesforce Agentforce 360

• Best for: enterprises already on Salesforce + Slack; predefined GTM/Service workflows; enterprise IT alignment.
• Interoperability: reasoning model choice (OpenAI/Anthropic/Google) and Slack integration road‑map. TechCrunch.
• Consider: platform lock‑in to CRM stack; ensure external tool coverage via connectors.

Google Vertex AI Agent Builder / Agent Engine

• Best for: multi‑cloud teams seeking managed runtime, sandboxed code execution, memory bank, and A2A agent‑to‑agent flows.
• Interoperability: embraces A2A and OSS frameworks (LangGraph, CrewAI) with governance primitives. Product; Release notes.
• Consider: align costs for managed runtime + evals; train team on Agent Engine concepts.

Amazon Nova Act (research preview)

• Best for: browser‑automation patterns and developers exploring web‑control agents with an SDK.
• Interoperability: complements AWS data/infra strategy; watch for governance maturity.
• TechCrunch.

Notion Agents

• Best for: teams living in Notion who want agentic analysis and task automation on workspace data.
• TechCrunch.

Where Microsoft Agent 365 fits

Think of Agent 365 as the management and governance layer over your agent ecosystem—authorize, quarantine, audit, and measure agents across vendors (including Salesforce), not the place you build agents. Reuters.

A 14‑day bake‑off plan (bring your own use cases)

Goal: identify a primary build platform and a governance stack you can run in production, safely, before year‑end.

1. Days 1–2: Define success. Pick 2–3 measurable use cases (e.g., customer support auto‑reply, SEO brief generation, agentic checkout handoff). Write acceptance criteria (task success rate, time to resolution, human‑in‑the‑loop thresholds, PCI/PII rules). See our SEO Agent in 7 Days and Agentic Checkout plan.
2. Days 3–4: Stand up governance. Create an agent registry + RBAC, logging, and guardrails. If you’re a Microsoft shop, map supervision to Agent 365 scopes.
3. Days 5–7: Build thin slices on each platform. Implement the same use case on 2–3 contenders (AgentKit, Agentforce 360, Vertex). Keep integrations identical (same RAG index, same tools, same channels). Document build effort.
4. Days 8–10: Run evals + load tests. Capture task completion, latency, cost per successful task, and fallbacks. Add observability (traces + step logs). Use MCP/A2A where available to keep swaps cheap. OpenAI tools, Vertex Agent Engine.
5. Days 11–12: Security & compliance. Validate OAuth/OIDC flows, data residency, PII/PCI scopes, and approval workflows. Keep a change log in your registry.
6. Days 13–14: Decide + rollout. Choose 1 primary platform + 1 backup. Publish an internal SOP, SLAs, escalation runbooks, and KPIs. Connect to Agent 365 (if applicable) for fleet oversight.

Copy‑paste RFP snippet (edit for your company)

Section 1: Scope & Use Cases
- Channels: web chat, email, helpdesk, voice (SIP optional)
- Data: existing RAG index (vector + keyword), knowledge bases
- Tools: order status API, CRM, CMS, payments (tokenized)

Section 2: Security & Governance
- Agent identity model, OAuth/OIDC, secret isolation
- Registry, approvals, audit logs, incident response
- Observability: traces, step logs, eval harness, red‑flag alerts

Section 3: Interoperability
- Support for MCP servers and/or A2A protocol
- Import/export of agent specs and policies

Section 4: Performance & Cost
- Task success rate at P95 latency target
- Cost per successful task, rate limits, autoscaling behavior

Section 5: References
- Production case studies in similar verticals
- Compliance attestations (SOC 2, PCI scope notes)

Recommendations by scenario

• Early‑stage SaaS: Start with AgentKit for speed, layer a light registry/RBAC, and keep MCP connectors portable. Add Agent 365 later if you’re already on Microsoft 365.
• Salesforce‑centric enterprise: Pilot Agentforce 360 + Slack. Ensure external tool coverage and exportability of agent specs.
• Data‑heavy, multi‑cloud teams: Evaluate Vertex Agent Builder/Engine for A2A flows, sandboxed code execution, and managed runtime.
• Retail/e‑commerce checkout: Prioritize governance patterns and PCI/SCA mapping; keep agents discoverable via your registry and route oversight into Agent 365 if you’re on Microsoft. See our PCI + SCA guide.

Key takeaways

• Pick the platform that fits your workflow and governance today, but design for swap‑ability tomorrow via MCP/A2A.
• Measure cost per successful task, not just model pricing.
• Treat Agent 365 (or equivalent) as your control plane; treat the build platform as interchangeable.