Research brief

Status: complete; published 2026-08-19 Scope finalized: 2026-08-19, after maintainer POV (kickoff questions 4-6) Knowledge input: maintainer kickoff answers (estate composition, gateway-first MCP stance, four nominated debates)

Scope (finalized inventory)

• Current estate: SOA and ESB heritage, API management, EDI and B2B, iPaaS, event backbone, workflow and BPM engines, RPA estate, legacy adapters (mainframe, file and batch)
• Agent-era: MCP tool layer and MCP gateways (distinct from LLM gateways), tool registries and connector catalogs, agent-to-agent protocols and cross-enterprise federation, the RPA-to-agent migration path
• Safe action: OAuth on-behalf-of and token exchange, idempotency and compensation patterns for agent-initiated actions, human-approval workflow integration, rate limiting
• Patterns: experience/process/system API layering, event-driven agent triggers, durable long-running orchestration

Layer-specific questions (agreed at kickoff)

1. Is the four-part composite estate (load-bearing ESB legacy, partial API coverage, iPaaS and RPA sprawl, underused event backbone) the accurate current-state picture? (H16)
2. Is MCP ready to be the enterprise tool contract now, and under what controls? (H17: gateway-first)
3. What do the auth standards arriving around MCP (EMA/XAA, ID-JAG, token exchange) change about the gateway's role?
4. How do agent-initiated writes stay safe (idempotency, compensation, approvals)?

Challenged defaults (nominated by maintainer, argued in findings)

• CD-2: dedicated MCP gateway products vs extending the API management layer
• CD-3: RPA estate: urgent replacement vs coexistence with agents supervising bots
• CD-4: agent-to-agent protocols now vs plain APIs until standards settle
• CD-5: legacy ESB: modernize as prerequisite vs bypass with a tool layer over systems of record

---

Source: research/R03-integration-fabric/brief.md in the evidence repository behind this site.