AI Agents

Salesforce's autonomous AI agent platform, built on the Atlas Reasoning Engine. Designed for enterprises that want agents grounded in CRM data, customer history, and Data Cloud unstructured sources. Every Atlas decision — planning, tool selection, action, reflection — is logged for audit, debugging, and behavior tuning.

• Atlas Reasoning Engine: plan → act → evaluate loop with full step-level audit logging
• Agentforce Script: hybrid agents combining deterministic workflows with LLM reasoning
• Agentforce Voice: AI voice agents across phone, web, and mobile channels
• Intelligent Context: low-code pipeline for unstructured/multimodal data grounding
• Native to Salesforce CRM, Data Cloud, Slack, MuleSoft — pre-wired to enterprise data
• Six pricing routes: Foundations (free starter), $2/conversation, Flex Credits ($500/100K = ~$0.10/action), per-user Add-ons (unlimited usage), Agentforce 1 Editions ($550/user/mo), and Service Cloud bundles

Limitations: Sticker shock at the high end — full Agentforce 1 Editions run $550/user/mo and complete deployments commonly land in $125–$650/user/mo territory. Best ROI sits inside the Salesforce ecosystem; outside it, the value proposition narrows. Implementation typically requires Salesforce Architects or partner help. Pricing model variety is flexible but adds forecasting complexity.

Microsoft's low-code platform for building, deploying, and managing AI agents across Microsoft 365, Teams, Power Platform, and standalone channels. Drag-and-drop topic authoring with code escape hatches; pre-built connectors to Dataverse, SharePoint, Dynamics, and 1,500+ Power Platform connectors.

• Multi-model: built-in Microsoft models, plus Azure Foundry BYO (GPT, Claude, Gemini, open-weight)
• Pricing: Copilot Credit packs at $200 / 25,000 credits / month, or pay-as-you-go meter
• Copilot Credit Pre-Purchase Plan for predictable enterprise spend
• Native integration with Microsoft 365 Copilot, Teams, Outlook, SharePoint, Dynamics 365
• Copilot Tuning, Copilot Connectors, agent governance via Microsoft Purview
• Strong enterprise posture: tenant isolation, EU Data Boundary, sovereign cloud options

Limitations: Credit consumption is opaque — specific actions have varying credit costs that aren't always obvious in advance. BYO Foundry models bill separately from Copilot Credits, complicating cost forecasting. The deepest value lives inside Microsoft 365; outside that ecosystem, you're paying for integrations you may not use. Heavy reliance on Power Platform conventions can feel limiting to code-first developers.

Google's two-pronged agent strategy. Antigravity (launched November 2025) is the agent-first IDE built around a "Manager Surface" for spawning multiple parallel coding agents with built-in browser self-verification. Vertex AI Agent Builder / Agent Engine is the production runtime for enterprise agents on Google Cloud, with managed sessions, memory, tool governance, and code execution.

• Models: Gemini 3 / 3.1 Pro by default; Claude Sonnet 4.6, Claude Opus 4.6/4.7, GPT-OSS 120B in Antigravity; Vertex Model Garden adds Anthropic + Mistral + open-weight
• Antigravity: Manager Surface for parallel agents, browser preview with verifiable artifacts, MCP support (added early 2026)
• Vertex Agent Engine: managed runtime with sessions, memory, tool governance; $0.0864 / vCPU-hr + $0.0090 / GB-hr for code interpreter
• Antigravity pricing: free public preview (rate-limited), AI Pro $20/mo, AI Ultra $249.99/mo
• Vertex pricing: token-based (matches Gemini API rates), pay-as-you-go on GCP
• Enterprise: VPC-SC, CMEK, IAM, audit logging, regional endpoints (Vertex side)

Limitations: Two distinct products with different audiences — Antigravity is coding-IDE-focused and still in public preview, while Vertex Agent Builder is the production answer but carries GCP setup and billing complexity. Antigravity rate limits have tightened post-launch and credit pricing on paid tiers is opaque. Vertex requires GCP comfort. The two surfaces don't yet share a unified developer story.

OpenAI's full agent toolkit, built on top of the Responses API. Agent Builder is a visual canvas with drag-and-drop nodes, guardrails, and version history. Agents SDK (Python & JS) is the code-first equivalent for multi-agent workflows. ChatKit embeds chat-based agent UIs into your product. Connector Registry centrally manages tools and data connections. Assistants API remains for legacy workloads but Agent Builder is the recommended path for new work.

• Models: GPT-5.5, GPT-5, GPT-5 mini, GPT-5-Codex, o-series reasoning models
• Agent Builder: visual canvas with versioning, eval configuration, preview runs
• Agents SDK: lightweight multi-agent framework (open source), handoffs, tracing
• ChatKit: drop-in chat UI components for embedding agents in your product
• Connector Registry: admin-managed tool/data connections across OpenAI products
• Pricing: token-based at model rates (no separate API fees); tool meters: Code Interpreter $0.03/session, File Search $0.10/GB/day, web search included in tool budget
• Built-in evaluation: datasets, trace grading, automated prompt optimization, third-party model support

Limitations: Tightest fit when you're already on OpenAI models — multi-vendor orchestration requires you to wire it up yourself. Agent Builder visual canvas is newer than enterprise alternatives like Copilot Studio. Pricing is straightforward at the API level but tool meters and storage can add up on large file-search workloads. Assistants API is being deprioritized in favor of the Responses-API-based stack — new builds should default to AgentKit.

The dominant open-source agent stack. LangGraph (MIT license, free) is a low-level graph-based orchestration framework for stateful, multi-actor agents. LangSmith is the paid observability and deployment layer — tracing, evals, and managed agent hosting. LangSmith Deployment (formerly LangGraph Platform, renamed October 2025) runs your agents at scale.

• Provider-agnostic: works with Anthropic, OpenAI, Google, Mistral, open-weight, anything with an API
• LangGraph: stateful graphs, single/multi/hierarchical agents, human-in-the-loop checkpoints, native streaming, built-in memory
• LangSmith Developer: free, 5,000 traces/mo, 14-day retention, 1 seat
• LangSmith Plus: $39/seat/mo, 10,000 traces, $2.50/1K overage, custom dashboards, evals
• LangSmith Enterprise: custom pricing, SSO, dedicated support, custom retention
• LangSmith Deployment: $0.005 per Deployment Run, custom for high volume
• Largest open-source agent ecosystem — thousands of integrations, templates, and community tools

Limitations: Code-first only — no drag-and-drop builder. Steeper learning curve than the managed platforms; you assemble the pieces yourself. LangChain (the broader framework) has historical reputation issues for over-abstraction; LangGraph is the cleaner, more focused successor. Self-hosting requires real infrastructure work; LangSmith Deployment exists exactly to avoid this. No native enterprise CRM/M365 integrations — you build them.

Best for: Enterprises whose IT, HR, customer service, or shared services backbone runs on ServiceNow and want autonomous agents that reason over the existing CMDB, knowledge bases, and process records — not just chat with end users. Strongest case where ticket volume is high and process patterns are stable.

• Thousands of pre-configured agents shipped across ITSM, HRSD, CSM, Field Service, Finance, Legal, Procurement, and industry verticals
• AI Agent Studio (low-code / natural language) for custom agents; Build Agent for IDE-out development
• Plan / act / reflect orchestration grounded in Workflow Data Fabric and Context Engine
• AI Control Tower for centralized governance, observability, and cost attribution
• BYO LLM on Prime tier (April 2026 retiering); Now LLMs and partner models on lower tiers
• Industry agent libraries: card disputes, insurance claims, telecom service activation, public-sector case workflows

Limitations: Agent quality is heavily dependent on Workflow Data Fabric and Context Engine grounding — without a clean data layer and modeled policies, agents hallucinate or refuse to answer. Biggest payoff when ServiceNow is already the system of record for the workflow being automated; less compelling for orgs running ServiceNow only as an ITSM tool. Pricing remains NDA — included in Foundation / Advanced / Prime tiers but token-pool overage can compound quickly.