Skip to main content

Chapter 6 · The Platform Wars: Who Is Building the Agentic Infrastructure

The decisions being made by platform providers today will constrain your options for years.


Why the Platform Layer Matters

Every agentic deployment sits on top of infrastructure it did not build: model APIs, orchestration frameworks, tool registries, memory systems, evaluation suites. The organisations that own this infrastructure are making architectural decisions — about interoperability, pricing, data access, and safety guardrails — that directly shape what you can and cannot build.

This is not new. Cloud platforms have always imposed constraints on what their tenants can do. What is different about the agentic infrastructure layer is the speed at which it is evolving, the degree to which early choices create lock-in, and the fact that multiple incompatible paradigms are competing simultaneously for dominance.

Understanding who is building what — and what their incentives are — is not background knowledge. It is a prerequisite for making sound architectural decisions.

One practical challenge with any treatment of this market is that it moves fast enough to outpace publication timelines. This chapter addresses that directly by separating two types of content: the analytical framework — four-layer model, competitive dynamics, lock-in logic — which is written for durability and should remain useful across multi-year timescales; and a Market Snapshot section at the end, which is explicitly dated and designed to be refreshed as the landscape changes. Read the framework for structural understanding; read the snapshot as a verified point-in-time reference.


The Four Layers of the Agentic Stack

The agentic infrastructure market stratifies into four distinct layers, each with different competitive dynamics.

Each layer shapes the one above it. Changes in foundation model capabilities propagate upward into orchestration frameworks. Changes in cloud infrastructure pricing propagate into which models teams can afford to run. Changes in tool standards determine how easily agents can connect to external systems. The structure of these layers maps directly onto the three fundamental capabilities any autonomous agent requires: planning, memory, and tool access — a taxonomy first comprehensively articulated by Weng (2023).3

By late 2025, the lower layers had begun to consolidate around clear winners — particularly at the standards layer, where both MCP and A2A moved to open, neutral governance under the Linux Foundation — while the orchestration framework and foundation model layers remained genuinely contested terrain.

A note on the diagram: the four layers are an analytical tool, not a fixed market structure. By 2025–26, foundation model providers had begun integrating downward into the infrastructure layer — Anthropic's Managed Agents service and OpenAI's Responses API with hosted tool execution both blur the line between Layer 4 and Layer 2. When a model provider also manages the runtime your agents run on, lock-in compounds across layers simultaneously. This is not a reason to avoid these services, but it is a reason to account for the compounded dependency.


Foundation Model Providers

OpenAI

OpenAI's strategic position is built on distribution. GPT-4 and its successors remain the most widely used models in enterprise deployments, not primarily because they are always the most capable on a given benchmark, but because they were first, have the broadest ecosystem, and are tightly integrated into the tools developers already use. The Assistants API introduced persistent threads and native tool use; the function calling specification became a de facto standard that competitors have largely adopted.

OpenAI's agentic ambitions are increasingly platform-oriented. The Operator initiative — agents that can navigate the web and interact with third-party services on a user's behalf — represents a direct bid to own the agentic application layer, not just the model layer.

Anthropic

Anthropic has differentiated on two dimensions: safety and coding. Claude's performance on software engineering benchmarks (SWE-bench) has driven strong adoption among developers. The Model Context Protocol (MCP) — a proposed open standard for connecting agents to tools and data sources — is the most significant infrastructure contribution any model provider has made to the ecosystem so far. If MCP achieves broad adoption, it reduces switching costs between models and shifts competition upward to capability and reliability. At launch, early integrations from companies including Block, Apollo, Zed, Replit, Codeium, and Sourcegraph indicated meaningful developer traction from the outset.1

Anthropic's enterprise focus is explicit. The computer use capability — allowing agents to operate a graphical interface — opens use cases that API-only models cannot address.

Anthropic's ambitions have also expanded beyond the model layer. Claude Managed Agents provides a fully managed agent harness — including container environments, tool execution, file access, web browsing, code execution, and session management — enabling developers to delegate agent orchestration to Anthropic's infrastructure rather than building their own runtime.13 This positions Anthropic as a direct competitor to the managed agent services of AWS Bedrock, Azure AI Foundry, and GCP Vertex AI, and introduces the compounded lock-in dynamic noted above: choosing Managed Agents ties you to both Claude and Anthropic's infrastructure simultaneously.

Google DeepMind

Google's advantage is integration depth. Gemini is embedded across Google Workspace, Search, and Cloud in ways that third-party models cannot replicate. For organisations already running significant workloads on GCP, the path-of-least-resistance for agentic deployment often runs through Vertex AI and Gemini.

Google's Agent Development Kit (ADK) and Agent-to-Agent (A2A) protocol represent a significant bet on agent interoperability — A2A designed as an open, cross-vendor standard built on existing web protocols rather than a Google-proprietary approach.6 Unlike most vendor frameworks, ADK was designed to be model-agnostic from the outset — supporting non-Google models via LiteLLM alongside native Gemini integration — and was already powering production deployments inside Google products at the time of its open-source release.2

Meta AI

Meta's contribution is structural rather than commercial. The Llama series — open-weight models that anyone can download, fine-tune, and deploy without API dependency — has fundamentally changed the competitive dynamics of the foundation model market. Llama 4's Mixture-of-Experts architecture delivers frontier-competitive performance at dramatically lower inference cost, and its open licensing enables deployment patterns (on-premise, air-gapped, fine-tuned) that proprietary models cannot match.

The Broader Competitive Tier

The foundation model landscape is not Western by nature, and treating it as such produces strategic blind spots. DeepSeek, the Chinese research lab, published V3 in December 2024 and R1 in January 2025, demonstrating that frontier-competitive reasoning performance could be achieved at dramatically lower inference cost than Western labs had assumed possible.14 The market reaction was immediate: every major Western provider's pricing roadmap was reassessed. DeepSeek is not merely a cost signal — it is evidence that the capital-intensive training assumptions underpinning Western model economics are not laws of physics. Alibaba's Qwen series has become a genuine global alternative to Meta's Llama in the open-weight tier, widely deployed in Western developer infrastructure for cost-sensitive workloads.15 Mistral AI, the French open-weight provider, serves organisations with European data sovereignty requirements as the leading independent foundation model company outside the US.

One dimension the four-layer framework must accommodate is geopolitical: Chinese-origin models, however technically capable, face regulatory scrutiny and procurement constraints in certain Western jurisdictions — particularly for deployments handling sensitive customer data, financial records, or government information. This is not a reason to exclude them from consideration, but it is a real constraint that belongs in any honest platform selection process.


Orchestration Frameworks

The orchestration framework market, chaotic and fragmented in 2023–24, had by late 2025 stratified into more distinct tiers. Code-first developer frameworks consolidated around a small set of production-proven options. In parallel, a second tier of visual and low-code platforms emerged to serve teams without deep software engineering capacity — a different buyer profile, a different risk model, and a different build-vs-buy calculus. (The decision between these tiers is developed further in Chapter 8.)

Tier 1 — Code-First Developer Frameworks

These frameworks are designed for teams building and operating agent systems in code, offering maximum control over state, memory, execution flow, and observability.

LangChain / LangGraph was the first widely adopted abstraction layer for building LLM applications, and its graph-based successor LangGraph has become the de facto framework for stateful multi-agent workflows in production. Its broad ecosystem of integrations and large community make it the default starting point for many teams. The trade-off: abstraction that simplifies building also obscures what is actually happening, which complicates debugging. LangGraph was created directly in response to that criticism — built for the control and production-readiness that the original framework lacked, and validated by enterprise deployments at LinkedIn, Uber, J.P. Morgan, and BlackRock.4

Microsoft Agent Framework is the product of Microsoft's consolidation of AutoGen and Semantic Kernel into a unified platform in October 2025, combining AutoGen's multi-agent conversation patterns with Semantic Kernel's plugin-based enterprise architecture.8 The merged framework supports multiple languages (C#, Python, Java), integrates natively with Azure, and offers the formal compliance and support guarantees that enterprise procurement requires. For teams in the Microsoft ecosystem, it is the natural consolidation point. In empirical evaluations conducted on the underlying AutoGen system prior to the merger, the multi-agent approach outperformed standalone GPT-4 and all compared commercial alternatives on standardised mathematical problem-solving tasks, with demonstrated enterprise applications including supply chain optimisation where it reduced implementation code by more than 75%.5

Key takeaway: AutoGen treats multi-agent coordination as structured conversation, and empirical benchmarks show this yields measurable performance gains over single-model approaches.

CrewAI takes a role-based approach — each agent in a crew has an explicit role, goal, and backstory — making it easier to reason about what each agent is supposed to do and to communicate agent designs to non-technical stakeholders. It has gained significant enterprise traction: the platform reports executing more than ten million agents per month, claims adoption across a significant portion of the Fortune 500, and raised $18M in Series A funding led by Insight Partners with Andrew Ng among its angel investors.10 A practical limitation noted by practitioners: teams commonly report hitting scaling walls after six to twelve months as requirements outgrow CrewAI's opinionated design, often requiring migration to LangGraph.

Provider-native SDKs — OpenAI Agents SDK, Google ADK (discussed above), and Anthropic's Agent SDK — each offer the tightest integration with their respective model ecosystems. The OpenAI Agents SDK deliberately occupies a minimalist position, built around four primitives (agents, handoffs, guardrails, and tracing), leaving advanced orchestration to the developer.12 These are best suited for teams already committed to a single provider; the lock-in trade-off is explicit by design.

Tier 2 — Visual and Low-Code Platforms

This tier serves a different buyer: technical and semi-technical teams who need to orchestrate agents across business workflows without the overhead of maintaining a full software engineering stack. The capability ceiling is lower, but deployment speed and organisational accessibility are substantially higher.

n8n is the leading platform in this tier. Fair-code licensed and self-hostable, it combines a visual workflow builder with the ability to write JavaScript or Python at any node — giving teams both speed and escape hatches. It supports 400+ integrations, native multi-agent and RAG patterns, and MCP tool calling, and its on-premise deployment option makes it viable in regulated environments where cloud-hosted orchestration is not.11 Unlike LangGraph, which demands software engineering expertise, n8n can be operated by technically proficient business analysts, making it the primary entry point for agentic workflows in organisations without dedicated AI engineering teams.

Dify, with 129k GitHub stars, occupies similar ground with a stronger emphasis on visual drag-and-drop development, making it accessible to non-developers building simpler agent workflows.

Data and Retrieval

LlamaIndex focuses on data — specifically, on connecting agents to structured and unstructured data sources. Its retrieval engine and query pipeline architecture are well suited to knowledge-intensive agentic applications where the quality of information retrieval is the primary performance lever.


Cloud Infrastructure Providers

The three major cloud providers have each built managed services for agentic AI, and each reflects its parent company's strategic interests. As noted above, the line between foundation model providers and cloud infrastructure providers is no longer clean: Anthropic now competes directly in this layer.

PlatformKey OfferingDifferentiationBest Fit
AWS BedrockManaged agent runtime, multi-model access; AgentCore GA Oct 2025AWS ecosystem, enterprise complianceTeams already on AWS, compliance-heavy industries
Azure AI FoundryCopilot Studio, Microsoft Agent Framework integrationMicrosoft 365 and Azure integrationOrganisations in the Microsoft ecosystem
GCP Vertex AIAgent Builder, Gemini native, ADK-native deploymentGoogle data services, multimodalData-intensive workloads, GCP-native teams
Anthropic Managed AgentsFully managed Claude agent harness; containers, tool execution, MCP, sessionsTightest Claude integration; vertical stack from model to runtimeTeams building Claude-native systems; avoids self-managed infrastructure13

Each platform locks in at the infrastructure layer rather than the model layer — you can often swap models while staying within the platform. The switching cost is in the tooling, monitoring, identity management, and compliance infrastructure you build on top. For Anthropic Managed Agents, the switching cost is compounded: both the model and the runtime would need to be replaced simultaneously.


The Standards Battle: MCP, A2A, and the Emerging Stack

The most consequential architectural question in the agentic ecosystem is how agents connect to tools and to each other. What began as a three-way contest has both partially resolved and expanded.

Function calling (OpenAI's original specification, now widely adopted) defines how a model can invoke a structured tool and receive a result. It is well understood and universally supported, but is a protocol for a single model–tool interaction, not for inter-agent communication or persistent tool registries.

Model Context Protocol (MCP) extends this — defining a standard for tool servers that any compatible agent can discover and invoke, regardless of which model is running. An MCP server exposes a set of capabilities; an MCP client (the agent) queries and uses them. This decouples the tool from the model, enabling a marketplace of reusable tools and reducing integration work significantly. At launch, early adopters including Block, Apollo, and Sourcegraph signalled meaningful traction.1 In December 2025, Anthropic donated MCP to the Agentic AI Foundation (AAIF) — a directed fund under the Linux Foundation, co-founded by Anthropic, Block, and OpenAI — establishing it as a vendor-neutral open standard.9

The significance of this governance shift is practical: it means no single company can modify the protocol to suit its commercial interests, the standard will outlive any one contributor's commercial fortunes, and organisations that previously hesitated to adopt a vendor-controlled standard can now do so without the political or financial risk that proprietary dependency creates. MCP can now be treated as infrastructure rather than a vendor bet — the same signal that Kubernetes moving to the CNCF sent to enterprise infrastructure teams in 2016.9 That governance signal should not be confused with automatic security maturity: MCP servers still execute tool calls inside real environments, so procurement teams should evaluate server provenance, update cadence, sandboxing, and permission boundaries with the same seriousness applied to any other integration layer.

Agent-to-Agent (A2A) is Google's proposed standard for agents communicating with each other — passing tasks, contexts, and results across agent boundaries. Where MCP addresses the agent-to-tool relationship, A2A addresses agent-to-agent coordination.6 At launch it carried contributions from more than 50 technology and services partners — including Salesforce, SAP, ServiceNow, and Workday — signalling unusually broad enterprise buy-in for a protocol still in draft specification.6 A2A was subsequently donated to the Linux Foundation in June 2025, applying the same neutral governance logic as MCP, and the same enterprise adoption signal follows:16 the risk of backing a vendor-controlled agent-communication protocol has been substantially reduced.

These standards are not mutually exclusive, and real deployments are likely to use combinations of all three. But the Linux Foundation governance for both MCP and A2A marks a meaningful moment: the lower layers of the standards stack are converging toward neutral ground, reducing the risk of betting on a vendor-controlled protocol. Cisco's AGNTCY initiative — introducing its own Agent Connect Protocol for cross-framework agent interoperability, backed by LangChain, LlamaIndex, and CrewAI — is early evidence that the contest above this layer is still broadening.7


What This Looks Like in Practice: The Same Workflow on Different Platforms

Platform choice does not merely change the developer experience. It changes what the organisation can control, how quickly it can move, what it can audit, and how much switching cost it accumulates. Public evidence rarely provides clean, apples-to-apples comparisons of the same enterprise agent implemented across multiple platforms. The available evidence is more modest: framework papers show that orchestration choices affect agent behaviour and performance, while vendor case studies show what particular platforms make easier in their own ecosystems. AutoGen, for example, demonstrates how multi-agent conversation patterns can combine LLMs, tools, and human inputs across varied applications.5 BOLAA provides comparative evidence that agent architecture and model backbone choices materially affect performance in autonomous-agent tasks.19 Salesforce Agentforce customer stories illustrate the enterprise-suite route: agents embedded in CRM and customer-facing workflows, with the usual caveat that vendor case studies should be read as deployment evidence, not neutral performance comparisons.20

Consider the same business problem: a customer-support agent that answers account questions, drafts responses, escalates exceptions, and updates CRM records only after approval. The implementation changes meaningfully depending on the platform layer chosen.

Platform routeWhat improvesWhat becomes harderBest fit
Managed provider agentFast setup, hosted tools, provider-managed runtimeLess control over orchestration internals; model/runtime lock-inTeams prioritising speed and low infrastructure burden
Code-first frameworkMaximum control over state, routing, evaluation, and observabilityHigher engineering and operations burdenComplex workflows with custom governance or domain logic
Enterprise suiteNative access to CRM, identity, case history, and business objectsDependency on platform roadmap and data modelSales, service, marketing, and customer operations already centred on that suite
Low-code workflow platformFast adoption by technical business users; easy integration with common toolsLower ceiling for complex stateful autonomy and custom evaluationDepartmental automation and Stage 2–3 workflows
FieldPractical design question
ProblemIs the goal to automate a standard workflow, or to encode differentiated business logic?
SetupDoes the agent live inside an enterprise suite, a code-first framework, or a managed provider runtime?
Agentic elementDoes the system choose tools, adapt steps, and escalate based on intermediate results — or only follow a fixed workflow?
Tools/dataWhich systems are native to the platform, and which require custom integration?
Human oversightCan approval, audit, rollback, and exception handling be expressed in the platform itself?
Main riskA platform that accelerates the pilot may later constrain portability, evaluation depth, or governance.

The point is not that one route is universally better. The point is that platform selection is an architectural decision disguised as a procurement decision. It determines not only how the agent is built, but also how it will be monitored, governed, scaled, and eventually replaced.


How to Think About Platform Lock-In

The agentic platform market is moving fast enough that decisions made today will likely need to be revisited. The practical implication is not to avoid lock-in at all costs — some lock-in is the price of moving quickly — but to be deliberate about which layer you allow yourself to depend on.

Lock in at the application layer, not the model or framework layer. The model that is best today will not be best in eighteen months. The framework that is popular today may be abandoned in two years. Build your business logic to be model-agnostic and framework-replaceable.

LayerLock-In RiskMitigation
Foundation modelMediumUse abstraction layers; maintain secondary model access
Orchestration framework — code-firstHighKeep framework-specific code thin; abstract core logic
Orchestration framework — visual/low-codeMediumWorkflow definitions are more portable than compiled code
Cloud infrastructureHighInfrastructure-as-code; avoid proprietary services where alternatives exist
Model + infrastructure (e.g. Managed Agents)Very HighDual dependency; requires deliberate justification
Tool integration standardsLow (MCP/A2A now Linux Foundation)Adopt open standards; MCP and A2A are now safe infrastructure bets

Market Snapshot · Q2 2026

This section reflects the platform landscape as of Q2 2026. The analytical framework above this line is written for durability; the tables below are a verified point-in-time view and should be treated accordingly. When significant market events occur — a major merger, a governance change, a significant new entrant — this section is the appropriate update target.

Foundation Models — Current Positioning

ProviderGeographyModel type2026 positioning
OpenAIUSProprietaryDistribution leader; GPT-4/5 family; Agents SDK; Responses API
AnthropicUSProprietarySafety + coding differentiation; MCP originator; Managed Agents infrastructure play
Google DeepMindUSProprietaryIntegration depth across Workspace, Search, Cloud; ADK; A2A
Meta AIUSOpen-weightLlama 4; enables on-premise and air-gapped deployments; structural market influence
Mistral AIEUOpen-weightEuropean sovereignty alternative; leading independent provider outside US
DeepSeekCNOpen-weightFrontier performance at low inference cost; disrupted Western pricing assumptions (Jan 2025)
Alibaba (Qwen)CNOpen-weightGlobal competitor to Llama; widely deployed in Western infrastructure

Note: Chinese-origin models face regulatory and procurement constraints in certain Western jurisdictions for sensitive workloads. Verify applicable restrictions before deployment.

Orchestration Frameworks — Production Signal

FrameworkTierKey evidenceBest for
LangGraphCode-first400+ companies in production; LinkedIn, Uber, J.P. Morgan, BlackRock; 90M+ monthly downloads; LangGraph 1.0 GA in October 202518Complex stateful workflows
Microsoft Agent FrameworkCode-firstGA April 2026; successor to AutoGen + Semantic Kernel; 10k+ orgs on Azure AI Foundry17Microsoft/Azure ecosystem
CrewAICode-first10M+ agent executions/month; $18M Series A; Fortune 500 adoptionRole-based multi-agent, rapid prototyping
OpenAI Agents SDKCode-firstLaunched March 2025; provider-agnostic; four primitivesTeams in OpenAI ecosystem
Google ADKCode-firstPowers Agentspace; 3.3M+ monthly downloads; model-agnosticGCP-native deployments
n8nVisual/low-code400+ integrations; self-hostable; MCP-compatibleTechnical teams without AI engineering capacity
DifyVisual/low-code129k+ GitHub starsNon-developer agent builders
LlamaIndexData/retrievalKnowledge-intensive RAG applications

Standards Governance Status

StandardStewardSinceEnterprise signal
MCPLinux Foundation (AAIF)Dec 9, 2025Settled — 97M monthly SDK downloads; 10k+ servers; safe infrastructure bet
A2ALinux Foundation2025Converging — 150+ supporting organisations; neutral governance confirmed
Function CallingOpenAI (de facto)2023Universal baseline — supported by all providers
ACP / AGNTCYCisco-led collective2025Contested — meaningful enterprise backing but not yet dominant

Notable Events Since Q2 2025

  • Oct 1, 2025 — Microsoft launched Microsoft Agent Framework (public preview), merging AutoGen and Semantic Kernel. GA reached April 2026.17
  • Dec 9, 2025 — MCP donated to Linux Foundation Agentic AI Foundation (AAIF), co-founded by Anthropic, Block, and OpenAI.
  • Dec 2024 / Jan 2025 — DeepSeek V3 and R1 released, disrupting Western model pricing assumptions with frontier-competitive performance at substantially lower inference cost.
  • March 11, 2025 — OpenAI launched Agents SDK alongside Responses API as production successor to the experimental Swarm framework.
  • 2025 — Anthropic launched Claude Managed Agents, entering the managed infrastructure layer alongside AWS Bedrock AgentCore, Azure AI Foundry, and GCP Vertex AI.
  • Oct 2025 — LangGraph 1.0 released as a stable production framework milestone.18
  • Oct 2024 — CrewAI raised $18M Series A (Insight Partners, Andrew Ng); $44.5M total to date.
  • 2025 — A2A donated to Linux Foundation, reaching 150+ supporting organisations.


This chapter maps the external platform landscape: the models, frameworks, clouds, and standards that shape what agent systems can become. Chapter 7 turns from platform selection to operating mode. Once agents can be deployed on this infrastructure, the next frontier is not only what they can do, but whether they remain continuously present — monitoring, remembering, and acting in the background.

References

  1. Anthropic (2024). Introducing the Model Context Protocol. Anthropic. https://www.anthropic.com/news/model-context-protocol
  2. Huizenga, E. & Yang, B. (2025). Agent Development Kit: Making it easy to build multi-agent applications. Google Developers Blog. https://developers.googleblog.com/en/agent-development-kit-easy-to-build-multi-agent-applications/
  3. Weng, L. (2023). LLM Powered Autonomous Agents. Lil'Log. https://lilianweng.github.io/posts/2023-06-23-agent/
  4. Chase, H. (2025). Reflections on Three Years of Building LangChain. LangChain Inc. https://www.langchain.com/blog/three-years-langchain
  5. Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., Jiang, L., Zhang, X., Zhang, S., Liu, J., Awadallah, A.H., White, R.W., Burger, D., & Wang, C. (2023). AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. arXiv:2308.08155. Microsoft Research.
  6. Surapaneni, R., Jha, M., Vakoc, M., & Segal, T. (2025). Announcing the Agent2Agent Protocol (A2A). Google Developers Blog. https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/
  7. Menon, P. (2025). Beyond MCP and A2A: How Cisco is Building the Internet of Agents. Cisco Live 2025, Session PSOETI-1113. Outshift by Cisco.
  8. Azure AI Foundry Team (2025). Introducing Microsoft Agent Framework: The Open-Source Engine for Agentic AI Apps. Microsoft Foundry Blog, October 1, 2025. https://devblogs.microsoft.com/foundry/introducing-microsoft-agent-framework-the-open-source-engine-for-agentic-ai-apps/
  9. Anthropic (2025). Donating the Model Context Protocol and establishing the Agentic AI Foundation. Anthropic, December 9, 2025. https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation
  10. Insight Partners (2024). How CrewAI is orchestrating the next generation of AI Agents. Insight Partners. https://www.insightpartners.com/ideas/crewai-scaleup-ai-story/
  11. n8n (2025). AI Agents in n8n. n8n.io. https://n8n.io/ai-agents/
  12. OpenAI (2025). New tools for building agents. OpenAI, March 11, 2025. https://openai.com/index/new-tools-for-building-agents/
  13. Anthropic (2025). Claude Managed Agents overview. Anthropic API Documentation. https://platform.claude.com/docs/en/managed-agents/overview
  14. DeepSeek-AI (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv:2501.12948. January 2025. https://arxiv.org/abs/2501.12948. Also see DeepSeek-AI (2024). DeepSeek-V3 Technical Report. arXiv:2412.19437. December 2024. https://arxiv.org/abs/2412.19437
  15. Qwen Team (2024). Qwen2.5 Technical Report. arXiv:2412.15115. December 2024. https://arxiv.org/abs/2412.15115
  16. Linux Foundation (2025). Linux Foundation Launches the Agent2Agent Protocol Project to Enable Secure, Intelligent Communication Between AI Agents. June 23, 2025. https://www.linuxfoundation.org/press/linux-foundation-launches-the-agent2agent-protocol-project-to-enable-secure-intelligent-communication-between-ai-agents
  17. Microsoft Foundry Team (2026). What's new in Microsoft Foundry | April 2026. Microsoft Foundry Blog. April 2026. https://devblogs.microsoft.com/foundry/whats-new-in-microsoft-foundry-apr-2026/
  18. LangChain Team (2025). LangGraph 1.0 is now generally available. LangChain Changelog. October 22, 2025. https://changelog.langchain.com/announcements/langgraph-1-0-is-now-generally-available
  19. Liu, Z., Yao, W., Zhang, J., Xue, L., Heinecke, S., Murthy, R., Feng, Y., Chen, Z., Niebles, J.C., Arpit, D., Xu, R., Mui, P., Wang, H., Xiong, C., & Savarese, S. (2023). BOLAA: Benchmarking and Orchestrating LLM-Augmented Autonomous Agents. Salesforce AI Research. arXiv:2308.05960. https://arxiv.org/abs/2308.05960
  20. Salesforce (2026). Customer Stories. Salesforce. https://www.salesforce.com/customer-stories/

Building agentic AI and wondering why alignment is harder than the technology? Get in touch