// capabilities · platform engagements

Where we deploy and where we migrate away from.

The enterprise AI field has broadened — native SaaS platforms, open graph frameworks, and the new agent SDKs now coexist, with MCP as the shared tool-connection layer underneath. We have a stance on each. Vendor-neutrality is a sales position, not an engineering one.

Salesforce Agentforce

deploy here

Strong fit when CRM is the system of record and the agent's primary tools are inside the Salesforce surface. Native data access, native action gating, native audit trail.

// we accept
  • Customer service triage and case routing inside Service Cloud
  • Renewal workflows where Salesforce CPQ is the source of truth
  • Sales prospecting agents tightly bound to Sales Cloud data
// we avoid
  • Workflows that span Salesforce + 4+ external systems — the orchestration tax leaves the platform
  • Use cases requiring custom guardrail policies beyond the built-in Trust Layer

Microsoft Copilot Studio

deploy here

Strong fit inside the Microsoft 365 + Dynamics surface. We deploy here when the buyer is already standardized on M365 and the agent surface is Teams or Outlook.

// we accept
  • Productivity agents inside Teams (meeting prep, status digest, scheduling)
  • Internal helpdesk routed through Outlook and SharePoint
  • Dynamics-bound CRM and ERP automation
// we avoid
  • Public-facing customer agents — the brand surface and customization are weaker
  • High-volume API workloads — the platform's connector limits surface fast

Google Vertex AI Agent Builder

selective

Strong primitives, weaker product. We deploy on Vertex when the buyer is GCP-native and the workload is data-heavy (BigQuery as the source of truth). We avoid it for workflow-heavy agents where the orchestration story is still maturing.

// we accept
  • BigQuery-backed analytical agents (NL → SQL → result)
  • RAG over GCS-hosted document lakes
  • Workloads where Gemini's long context is a structural advantage
// we avoid
  • Multi-agent topologies — the framework story is behind LangGraph and the agent SDKs
  • Workflows requiring deep tool ecosystem (still maturing vs. M365 / Salesforce)

Agent SDKs — Claude / OpenAI / Vercel AI

deploy here

The standard libraries for building agents in 2026. Lighter than a full graph framework, with first-class MCP tool support, session + memory primitives, and streaming out of the box. We reach for these when the agent is one actor with tools, not a multi-node graph with routing policies.

// we accept
  • Single-actor agents with tool use — document agents, analysis assistants, support copilots
  • MCP-first tool surfaces that need to stay portable across model vendors
  • Streaming UIs where the Vercel AI SDK's React primitives eliminate boilerplate
  • Extended-thinking workloads where reasoning tokens and structured outputs are first-class
// we avoid
  • Complex multi-agent graphs with retry edges and reviewer nodes — LangGraph is still the right tool there
  • Workflows where the graph topology is the load-bearing abstraction

LangGraph / custom topologies

deploy here

Our default when the workflow spans multiple systems, needs custom guardrail nodes, or has compliance constraints the SDKs and SaaS platforms can't satisfy. The agent SDKs took the simple end of the spectrum; LangGraph is still where real graph-shaped work lives. CrewAI stays in the toolkit; AutoGen we use less now that Semantic Kernel absorbed most of its patterns.

// we accept
  • Multi-system orchestration across CRM, ERP, payments, and external APIs
  • Custom guardrail policies that need to live in version control
  • Regulated industries (insurance, healthcare, financial services)
  • Workflows requiring replayable traces in your observability stack
// decision grid

How we pick.

The questions we walk through on the first discovery call. The order matters — the answer to question one constrains every answer below it.

  1. Q01

    Is the system of record one of Salesforce / Dynamics / BigQuery?

    → Lean toward the native platform (Agentforce / Copilot / Vertex) for the primary integration.

  2. Q02

    Is the agent one actor with tools, or a graph of actors with routing policies?

    → One actor → the agent SDKs (Claude, OpenAI, Vercel AI). Graph with routing → LangGraph.

  3. Q03

    Are your tools MCP-standardized today, or still bespoke?

    → MCP where you can. It makes tool definitions portable across model vendors and lets us swap the LLM without rewiring the integration surface.

  4. Q04

    Do you need custom guardrails or replayable traces beyond what the platform offers?

    → Build on LangGraph with the SaaS platform as one tool among many; keep policies in version control and traces in your observability stack.

  5. Q05

    Is the workflow customer-facing with brand surface requirements?

    → Custom front-end + LangGraph or SDK backend. The SaaS platforms are weaker at brand control.

  6. Q06

    Is this regulated (insurance / healthcare / payments)?

    → Custom topology. Compliance teams need policies in version control and traces in their observability stack — not in a vendor's black box.

// next

Hybrid platform engagements in production.

See the workBook a session