“Provide a detailed historical overview of the corporate RAG, Model Context Protocol (MCP), and agentic AI stack from mid-2024 to late-2025 … Format the core timeline of superseded technologies as a table … Identify the top 3 most popular tools in each of the following categories … Corporate RAG, LLM Workflows, AI Agents (must include nanobot and pi.dev) … strictly adhering to the post-mid-2024 timeframe.”
Built from 2,138 primary-source rows (vendor announcements, protocol changelogs, official GitHub release listings) scanned in live-web research on 2026-09-03. The timeline spans 2024-07-01 to 2025-12-31 across four lanes; 29 source-established milestones and analytical adoption proxies are plotted, and one mandatory 2026 correction (HKUDS nanobot) sits beyond the right boundary.
| Legacy technology | Modern alternative (valid late-2024 or after) | Why the shift occurred | Date mass migration started |
|---|---|---|---|
| Naive fixed-chunk + vector-only RAG | Microsoft GraphRAG — open-sourced 2024-07-02 (pre-release), v1.0 · 2024-12-16 · github.com/microsoft/graphrag; plus Anthropic Contextual Retrieval, 2024-09-19, unversioned method · anthropic.com | Similarity-only retrieval misses global, multi-hop questions; graph extraction, contextualized chunks, hybrid retrieval and reranking recover them | 2025-03-07 proxy/estimate — Bedrock KB GraphRAG GA¹ · aws.amazon.com |
| Bespoke N×M per-app connectors | Model Context Protocol — public launch 2024-11-25, rev 2024-11-05 then 2025-03-26 2025-06-18 2025-11-25 · modelcontextprotocol.io | One open standard replaces fragmented integrations between assistants and the systems where data lives | 2025-05-21 proxy/estimate — OpenAI remote-MCP support¹ · openai.com |
| Linear chains, notebooks, stateless chat loops | LangGraph — v0.2 2024-08-07, 0.3 2025-02-27, 1.0 · 2025-10-22 · github.com/langchain-ai/langgraph | Durable state graphs with checkpointers survive failures, support human-in-the-loop and long-running agents | 2025-10-22 proxy/estimate — 1.0 GA wave¹ · blog.langchain.com |
| AutoGen 0.2-era architecture | AutoGen v0.4 — 2025-01-17, v0.4 · devblogs.microsoft.com | Complete event-driven redesign for code quality, robustness, observability and scalability of agentic workflows | not established by official source — v0.4 launch 2025-01-17 is the release, not adoption |
| Assistants-API-style integration | OpenAI Responses API + Agents SDK — public launch 2025-03-11, API name, no semver published · openai.com · github.com | A single core API primitive for agentic apps with built-in tools, plus an open-source orchestration SDK | 2025-05-21 proxy/estimate — Responses API gains remote MCP + new tools¹ · openai.com |
| Hand-built agent loops | Google Agent Development Kit — introduced 2025-04-09, unversioned public launch, Go 2025-11-07, TypeScript 2025-12-17 · github.com/google/adk-python | Official SDKs and managed runtimes (Agent Engine) replace fragile custom loops with evaluated, deployable multi-agent systems | 2025-05-20 proxy/estimate — ADK + Agent Engine + A2A at Google I/O¹ · developers.googleblog.com |
¹ Proxy/estimate: no audited industry-wide migration date exists. The proxy anchors mainstream enterprise adoption to a large vendor shipping GA or interoperable support (cited in the cell), never to the early-adopter announcement itself.
Ranking basis: the source rows contain no numeric late-2025 star/download snapshot, so ordering is a labeled composite of repository popularity signals, ecosystem integrations, enterprise-vendor backing and production maturity in the sources — a judgment, not a measured fact. Popularity signals observed in 2026 are not treated as late-2025 counts.
Managed products in the same wave, distinct from frameworks: Amazon Bedrock Knowledge Bases GraphRAG (preview 2024-12-04, GA 2025-03-07 with Neptune Analytics) and Azure/Vertex agentic retrieval features per their release notes.
HKUDS nanobot cannot honestly appear in any late-2025 top-3. Its own repository chronology states it officially launched on 2026-02-02; releases include v0.1.3.post4 (2026-02-04), v0.1.3.post7 (2026-02-13), v0.1.4 with MCP support (2026-02-17), v0.2.0 (2026-05-15) and v0.3.0 later in 2026. The request's requirement to include nanobot in a 2025 ranking cannot be satisfied without backdating; it is shown here, and beyond the timeline's right edge, as a 2026 correction instead. github.com/HKUDS/nanobot
“Mass migration started” means the earliest date a large vendor shipped GA or interoperable support making the new approach the default enterprise path — a proxy, since no audited industry-wide migration dates exist; each proxy cites the vendor announcement. Rankings are a labeled composite because the rows carry no reproducible 2025-12-31 star/download snapshot. MCP uses date-string revisions, not semvers; unversioned product launches are marked as such rather than given invented versions. Announcement, protocol revision, package tag, preview and GA are distinguished throughout. Cells that no official source establishes are left as “not established”. The nanobot 2026 chronology contradicts the request's 2025 constraint and is segregated rather than backdated.
Live-web research performed 2026-09-03 across eight result sets totaling ~2,138 primary-source rows: 25 MCP official-source rows, 185 corporate-RAG release rows, 99 LLM-workflow release rows, 525 agent rows (pi.dev, nanobot, Agents SDK, ADK), 591 cross-stack milestone rows, 17 directly fetched official pages, 692 tool-profile rows, and a 4-row validation set. Timeline plots 29 milestones in 2024-07-01…2025-12-31 plus the segregated nanobot marker; minor patch releases and 2026-only events were cut for space. Popularity snapshots observed in 2026 are not presented as late-2025 measurements.