28 Dec 2025 AI Engineering5 min read

2025 in Review: The Protocol Year

2025 was the protocol year: MCP became the cross-vendor tool interface, A2A moved to the Linux Foundation, reasoning became a commodity, and coding agents entered daily work. We put the verified milestones in context and identify the tasks for 2026: agent identity, evaluation infrastructure, and simulation.

The Interoperability Problem

In December 2024, almost every connection between a language model and an external system was custom code. Vendors brought their own function-calling formats, authentication paths, and frameworks. Connecting N models to M tools meant building and continually maintaining N×M integrations. The models were ready for real work; the infrastructure around them was not. That was how the year began.

In 2025, that infrastructure changed more than the models themselves. The defining events were specification revisions, foundation donations, registries, and adoption announcements — events at the protocol layer. That is why we call 2025 the protocol year. This retrospective records, with dates and numbers, what materially changed and states just as clearly what remained unsolved in October 2025.

Agents run in production — every run is a data point. 1/4

MCP Became the Default Interface

Anthropic released the Model Context Protocol on 25 November 2024 as an open standard for connecting AI applications to tools and data. On 26 March 2025, OpenAI adopted it: support landed in the Agents SDK immediately, with the ChatGPT desktop app and Responses API announced to follow. Microsoft had added MCP to Copilot Studio the week before. Google DeepMind confirmed Gemini support in April. Within five months, every major lab had adopted a competitor's protocol. That does not happen for marketing reasons; it happens when the cost of fragmentation exceeds the cost of cooperation.

The specification matured alongside adoption. The 2025-06-18 revision classified MCP servers as OAuth 2.1 resource servers, added elicitation and structured tool output, and removed JSON-RPC batching. On 8 September 2025, the official MCP Registry launched in preview at registry.modelcontextprotocol.io — an open catalog and API for publicly available servers. Its scope matters: MCP standardizes transport, capability discovery, and authorization, but it does not certify build quality, maintenance, or safety. The registry makes servers discoverable; it does not vet them.

A2A Moved to Neutral Ground

Google announced the Agent2Agent protocol on 9 April 2025 with more than 50 technology partners. A2A complements MCP rather than competing with it: MCP connects one agent to its tools; A2A lets autonomous agents discover each other through Agent Cards, exchange messages, and coordinate long-running tasks across vendor and framework boundaries.

On 23 June 2025, at Open Source Summit North America, Google donated the A2A specification, SDKs, and developer tooling to the Linux Foundation. Amazon Web Services, Cisco, Microsoft, Salesforce, SAP, and ServiceNow formed the project alongside Google; more than 100 companies back the protocol. Neutral governance removes a real enterprise adoption blocker — no one wants to build on a single vendor's agent protocol. The honest caveat: production A2A deployments remain rare compared to MCP servers. The protocol is ready. Most multi-agent architectures are not.

Reasoning Became a Commodity

In December 2024, OpenAI's o1 was the only production-grade reasoning model, priced at $60 per million output tokens. On 20 January 2025, DeepSeek released R1 under the MIT license: o1-comparable results on math and code benchmarks at $2.19 per million output tokens — roughly 96% cheaper — with six distilled variants from 1.5B to 70B parameters released alongside.

The rest of the year confirmed the pattern. Anthropic shipped extended thinking in Claude 3.7 Sonnet in February. Google shipped Gemini 2.5 Pro in March. OpenAI shipped o3 in April and the Apache-2.0 gpt-oss models in August. Alibaba released Qwen3 under Apache 2.0 in April. Reinforcement-learned chain-of-thought went from competitive moat to free download in under nine months. Reasoning is no longer a differentiator; it is a line item.

Model	Released	Weights	Output price per 1M tokens
OpenAI o1	5 Dec 2024	closed	$60.00
DeepSeek-R1	20 Jan 2025	MIT	$2.19
Qwen3-235B	29 Apr 2025	Apache 2.0	self-hosted
gpt-oss-120b	5 Aug 2025	Apache 2.0	self-hosted

Agents Entered Production

Coding agents also moved beyond the demo in 2025 and became daily tools. Anthropic made Claude Code generally available on 22 May, OpenAI launched Codex as a cloud software-engineering agent in May, and GitHub shipped a Copilot coding agent that takes assigned issues and opens pull requests. Long-running, multi-step, terminal-native agents thereby became a regular part of software delivery.

The measured picture is more sober than the announcements. METR's randomized controlled trial, published 10 July 2025, observed 16 experienced open-source developers working on their own mature repositories with early-2025 tools. Developers forecast a 24% speedup. The measured result was the opposite: tasks took 19% longer with AI assistance. The gap between perceived and actual productivity is itself the finding.

Our reading: agents deliver where output can be verified cheaply — greenfield code, boilerplate, tests, parallelizable migrations. In complex legacy systems, verification cost can exceed generation savings. Production agent value is real, but it is measured per workflow, not assumed per tool.

What the Protocols Do Not Solve

Identity is unsolved. MCP's OAuth 2.1 flow authenticates a client to a single server. There is no standard way to express that agent X acts on behalf of user Y with delegated scope Z, and no standard way to propagate that delegation across a chain of agents calling agents. A2A Agent Cards declare what an agent can do, not whether it should be trusted to do it.

Evaluation remains unsolved as well. Public benchmarks saturate within months and predict little about a specific agent in its actual workflow. Many teams therefore still discover regressions only in production — on their customers. Prompt injection through tool results also remains an open attack class: a protocol standardizes the pipe and, with it, the attack surface. None of the 2025 specifications closes these gaps; the protocols were not designed to do so.

What 2026 Must Deliver

Three expectations follow, written on 19 October 2025. First, agent identity becomes a standards problem. We expect work on delegation chains in the OAuth and IETF communities, alongside workload identity systems extended to agents. Whoever delivers scoped, auditable, and revocable agent credentials will unlock the enterprise deployments that this year's protocols support only on paper.

Second, evals become part of CI: trace-based regression suites will gate agent changes just as tests gate code changes today. Without that control, teams cannot ship agent updates responsibly. Third, simulation becomes standard practice. Before an agent touches a production system, it will have run thousands of episodes against a sandboxed replica. At Blue IT Systems, we treat all three as engineering disciplines, not research questions. In 2025, the industry standardized how agents speak; in 2026, it must establish how we test them and on what grounds we trust them.