About the Agentic History Museum
Last updated: June 1, 2026
Agentic History is the independent, unbiased authority on AI agent history, present developments, and future consequences.
Editorial source note: This page is the public record for how Agentic History chooses sources, handles contested firsts, accepts corrections, and separates primary evidence from museum interpretation. It supports the source notes used across the timeline, Primary Sources Library, AI Agent Taxonomy, Failure Archive, and Predictions Tracker.
Context
AI agents are not a 2023 invention — but the past three years have been the most consequential in the field's history. Foundational work on multi-agent coordination and agent architectures stretches back four decades. The recent wave of LLM-driven autonomous agents has pulled all of that lineage into mainstream attention, and is generating new primary materials at a rapid pace. This museum exists to document both — the deep history and the unfolding present — with care.
Our History page is the primary research output. The FAQ provides direct answers to the most common questions. The homepage summarizes the condensed timeline, key figures, and foundational papers. This page explains how the project is run.
Mission
- Document responsibly. Capture how autonomous agents are being built, deployed, and funded — as it unfolds — with verifiable sources.
- Preserve context. Link claims to dates, releases, papers, and public statements so readers can evaluate them in time.
- Surface lineage. Trace the line from the Contract Net Protocol and BDI architectures, through the multi-agent systems of the 1990s, to the LLM-driven systems of today.
- Support informed debate. Provide a neutral record that researchers, journalists, builders, and historians can cite, critique, and build upon.
- Counter revisionism. The history of a fast-moving field is often rewritten by those who profit from controlling the narrative. This museum maintains dated, sourced records precisely so that original claims can be compared against later accounts.
Scope & Inclusion Criteria
We include items that materially advance autonomous agents or shape their understanding:
- Primary sources: peer-reviewed papers, official product posts, lab announcements, release notes, conference talks, founders' own writing, and contemporaneous reporting.
- Substantive events: capability releases (planning, tool use, memory, computer use), open-source frameworks that materially shifted practice, notable funding rounds, and benchmark milestones.
- Comparative context: antecedents from multi-agent systems, distributed AI, BDI architectures, and reinforcement-learning agents when relevant to current practice.
We generally exclude pure marketing without verifiable details, unverifiable rumors, and minor product experiments unless they prove influential.
Research Methodology
Source Hierarchy
Primary Evidence Before Secondary Interpretation
- Sourcing. We prioritize first-party materials and high-quality reporting. Every substantive claim should be traceable to a public source.
- Dating. We record both the publication date and, when different, the event date. arXiv version dates are noted for academic papers. Revisions are logged with notes.
- Verification. Conflicting claims are noted side-by-side with links. When necessary, we add curator notes for clarity.
- Terminology. We use terms as the sources use them, noting when definitions differ. "AI agent," "autonomous agent," and "agentic AI" are not used interchangeably without comment.
- Funding / ecosystem. Rounds and valuations are listed with sources; we avoid speculative numbers and note when figures are reported but not officially confirmed.
- Maintenance. Pages carry a "last updated" stamp. Substantial edits appear in the changelog below.
Review Standard
What Gets Checked Before Publication
Before a claim is treated as part of the museum record, it should be checked for source quality, date accuracy, terminology fit, and relationship to existing pages. High-impact claims about safety, legal liability, money, or public harm require stronger evidence than ordinary descriptive claims.
Correction Standard
How Corrections Are Applied
Corrections should preserve the historical record rather than silently rewrite it. If a correction changes a date, source attribution, first-use claim, product classification, or verdict, the relevant page should be updated with clearer language and the change should be reflected through the page's last-updated date or changelog when material.
Neutrality on "Firsts"
"Firsts" in this field are unusually contested. Multiple projects in 2023 launched within days or hours of each other; multiple academic threads converged on similar ideas independently. Rather than adjudicate, we present original claims with dates and sources and mark them as contested where overlapping. Three places in our timeline call out specific contested origins:
- First publicly demonstrated end-to-end autonomous LLM agent. BabyAGI (March 28, 2023, Yohei Nakajima) and AutoGPT (March 30, 2023, Toran Bruce Richards) shipped within two days of each other. We treat them as effectively simultaneous and list both.
- First "computer use" agent from a frontier lab. Anthropic's Claude 3.5 Sonnet computer use (October 22, 2024) was the first widely deployed; OpenAI Operator (January 23, 2025) followed. Adept's ACT-1 (September 2022) demonstrated the underlying capability earlier but at smaller scale.
- "Pioneers of multi-agent systems." The honor is shared among Reid G. Smith (Contract Net Protocol, 1980), Marvin Minsky (Society of Mind, 1986), Yoav Shoham (Agent-Oriented Programming, 1993), and Stuart Russell & Peter Norvig (AI: A Modern Approach, 1995).
Readers can compare timelines in the History timeline and follow the citations to decide.
Editorial Positions on Key Historical Questions
Where the historical record is clear, we state it plainly. The following are our considered editorial positions on the most frequently contested questions in AI agent history. Each is held subject to correction if new primary sources emerge.
On the origin of modern LLM-based AI agents
The modern LLM-based AI agent has two identifiable origins that cannot be collapsed into one. The technical origin is the ReAct paper (Yao et al., arXiv October 6, 2022), which provided the architectural template — interleaved reasoning and action — that virtually all subsequent agent frameworks implement. The cultural and commercial origin is the BabyAGI/AutoGPT moment of March 2023, when autonomous LLM agents became publicly available and mainstream. The OpenAI ChatGPT API (March 1, 2023) is the enabling infrastructure event that made both possible at scale. These three events are each "firsts" in a different dimension; none supersedes the others.
On the depth of the historical lineage
The dominant industry narrative positions AI agents as a 2023 phenomenon. This is historically incorrect. The contract-net model (1980), BDI architecture (1987–1996), and multi-agent systems literature (1990s) described, formalized, and implemented many of the core ideas that LLM-agent builders rediscovered in 2022–2024 — often without knowing the prior literature. This is not a criticism of the modern pioneers; it is a statement that the intellectual history of the field runs decades deeper than most current accounts acknowledge. We document both the deep roots and the modern rediscovery.
On "who invented AI agents"
No single person or moment invented AI agents. The field accumulated, across decades, the components now recognized as defining: goal-directedness (Dartmouth, 1956), multi-agent coordination (Smith, 1980), intelligence-as-collective (Minsky, 1986), formal agent architecture (Bratman/BDI, 1987), software-engineering formalization (Shoham, 1993), educational standard (Russell & Norvig, 1995), LLM-as-agent-brain (ReAct, 2022), public availability (BabyAGI/AutoGPT, 2023). Any single-person "inventor" claim is a simplification that this museum resists.
On AgentSpeak(L) and Jason as underrepresented contributions
Dr. Anand Rao's AgentSpeak(L) (1996) and the Jason interpreter (Hübner and Bordini) represent a significant and historically underrepresented thread in the lineage from BDI theory to deployable agent software. The modern agent ecosystem largely rediscovered these ideas via the ReAct paper and LangChain without awareness of the prior work. We note this parallel development explicitly rather than presenting the 2022–2023 literature as an ex nihilo invention.
How to Contribute Primary Sources
We welcome corrections, missing primary sources, and oral histories from participants. If you submit a correction, please include:
- The exact passage or section to amend.
- A proposed correction with at least one public source link (first-party preferred).
- Any relevant dates (publication date vs. event date).
Email: curator@agentichistory.org.
Contact & Press Inquiries
For interviews, classroom use, or press inquiries, write to curator@agentichistory.org. We can provide a short overview, key dates, and a selected reading list tailored to your audience.
Changelog and Updates
- 2026-04-29: Initial public draft. Timeline covers 1956 through November 2025; funding table covers 2023–2026 representative rounds.
- 2026-05-04: Expanded About page with fuller mission statement, scope criteria, research methodology detail, contribution guidelines, and this changelog.
- 2026-05-07: Email received drawing our attention to Dr. Anand Rao's AgentSpeak(L) (1996) and its open-source interpreter Jason (Hübner and Bordini). Submission identified this as a significant and underrepresented link in the chain from BDI theory to deployable agent software.
- 2026-05-08: Curatorial review of AgentSpeak(L) and Jason submission completed. Primary sources verified: Rao (1996) MAAMAW-96 paper; Jason project documentation at jason-lang.github.io. Submission accepted.
- 2026-05-09: Timeline updated. The 1990s entry expanded to cover AgentSpeak(L) and Dr. Anand Rao's contribution. New dedicated timeline entry added for Jason (2005–ongoing). Two bibliography entries added.
- 2026-05-27: Major expansion of all pages. Homepage expanded with condensed 21-entry timeline, key figures section (foundational and modern pioneers), key papers section (11 primary sources), and five new definitional sections. FAQ expanded from 16 to 23 questions with substantially longer answers throughout; seven new questions added covering BabyAGI, MCP, LangChain, Devin, computer use, BDI architecture, and the Contract Net Protocol. About page expanded with new Editorial Positions section covering four contested historical questions. All pages updated to last-updated May 27, 2026.
Continue: Read the timeline · What is an AI agent? · AI Agent Taxonomy · Primary Sources Library · Failure Archive · Predictions Tracker · FAQ · Key Figures