Policy-governed agent orchestration
Controls for multi-agent workflows: where policy is evaluated, how approvals are expressed, and how decisions are recorded for audit.
Related guide →Research
Agentic integration research
Research areas, open problems and the standards shaping how AI agents connect to enterprise systems safely: MCP and A2A, governed APIs, agent identity, orchestration and auditability.
Research areas
Themes the guides on this site investigate, each linked to the most relevant guide.
Governed tool access through MCP gateways
Exposing enterprise APIs and integration flows as MCP tools with filtered discovery, per-tool authorization and stateless, header-routed traffic.
Related guide →Agent identity and delegated authorization
Carrying user, agent and service identity across every hop so an agent never exceeds the permissions of the person it acts for.
Related guide →Cross-platform agent interoperability
A2A delegation between agents on different platforms and organizations: trust in Agent Cards, task lifecycles and partner onboarding.
Related guide →Agentic risk controls in integration layers
Mapping agent-specific risks such as goal hijack, tool misuse and cascading failures to controls in gateways, API management and integration flows.
Related guide →Event-driven and long-running agent work
Triggering agents from business events and keeping long-running tasks consistent with idempotency, compensation and correlation.
Related guide →Open problems in 2026
Questions the protocols and standards do not yet fully answer, and where integration practice is still forming.
Multi-hop delegation
OAuth handles one hop well. Chains of agents across organizations and asynchronous pauses still lack a settled pattern for preserving the original user's authority and narrowing it at each hop.
Tool selection reliability
Models choose tools from names and descriptions. How to write, test and version tool descriptions so the right tool is chosen consistently is largely empirical today.
Approval fatigue
Human approval is the most common safeguard and the easiest to weaken. Open question: which actions truly need a person, and how to show approvers evidence rather than an agent's summary.
End-to-end agent observability
Tracing a request from an agent conversation through gateways, integration flows and systems of record, with OpenTelemetry context and without logging personal data.
Retries in a stateless protocol
With stateless MCP and unresumable streams, retries become routine. Making every agent-reachable write idempotent across legacy systems is an engineering gap in many landscapes.
Evidence for regulation
What logs, approvals and traceability records integration layers must produce so organizations can demonstrate oversight of high-risk AI uses such as HR decisions.
Standards and specifications tracker
Status as of October 2026. Check the linked source for the latest version.
| Standard or initiative | Status | Maintained by |
|---|---|---|
| Model Context Protocol | Specification revision 2026-07-28 (stateless core) | Agentic AI Foundation |
| Agent2Agent (A2A) | v1.0, patch v1.0.1 | Agentic AI Foundation |
| OWASP Top 10 for Agentic Applications | 2026 edition, published Dec 2025 | OWASP GenAI Security Project |
| AI Agent Standards Initiative | Launched Feb 2026 | NIST CAISI |
| Agent identity and authorization concept paper | Comments closed Apr 2026; practice guide to follow | NIST NCCoE |
| Agent identity drafts (OAuth delegation, WIMSE) | Individual drafts; not yet standards track | IETF |
| EU AI Act, Digital Omnibus | Annex III high-risk duties from 2 Dec 2027 | European Union |
Have research, a case study or a correction to share? Get in touch.