Industries / Government, education and community / Education / EdTech
Learning platforms
The software that schools, colleges and employers use to run courses online: course sites, enrollment, gradebooks, and the shared formats that let course content move from one system to another.
7
Milestones achieved, from 1960 to 2024.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1960
A computer runs lessons
In 1960 the University of Illinois launched PLATO, which the university calls the first generalized computer-assisted instruction system, on its ILLIAC I computer. D. Bitzer, P. Braunfeld and W. Lichtenberger described it in a December 1961 paper titled "PLATO: An Automatic Teaching Device". By the late 1970s PLATO served thousands of graphics terminals around the world from almost a dozen networked mainframe computers, by the university's account.
Credit
Donald L. Bitzer and colleagues, University of Illinois
1989
PLATO is sold off
Control Data Corporation obtained the rights to PLATO in 1976 and began selling it in 1978. Each system could serve up to 1,000 users at a time, and around 1985 more than 100 systems were in use worldwide; the University of Amsterdam alone spent $200,000 to connect in the late 1970s. With Control Data in financial trouble, it sold PLATO to TRO Learning in 1989, according to the historian Paul Ceruzzi.
LessonCeruzzi lists cost first among PLATO's problems, including the high cost of writing lessons for courses with a short life, and the model of one central mainframe serving many terminals began to look outdated in 1986, when the first local-area networks were installed.
Credit
Control Data Corporation
2001
Course packages that travel
SCORM is a set of rules for packaging an online course so that any learning management system (LMS), the software that hosts courses and records students' progress, can run it and track the results. The Advanced Distributed Learning initiative (ADL) issued a draft in January 2000, and version 1.2, released in October 2001, became the widely used one. The US Department of Defense, in its Instruction 1322.26, requires its training content to be delivered through SCORM or its successors, xAPI and cmi5.
LessonThe first edition of SCORM 2004 contained sequencing rules that were not fully implementable and was never deployed, and the SCORM vendor Rustici Software says LMSs still read the rules slightly differently, so a course that conforms on paper does not always run everywhere.
Credit
ADL, building on earlier specifications from the AICC and IMS
2002
A free, open LMS
Moodle 1.0 was released on 20 August 2002. Moodle is open source: anyone may download, change and run its code under the GNU General Public License. By Moodle's count of the sites that register with it, 147,566 sites in 236 countries run Moodle, with about 513 million user accounts.
Credit
the Moodle project
2009
Course software patent falls
Blackboard, which the court called the market leader in course management software, sued its main rival Desire2Learn over a patent on an internet-based education system, and a Texas jury found that Desire2Learn had infringed it. On 27 July 2009 the US Court of Appeals for the Federal Circuit held the patent's asserted claims invalid, finding the key claim anticipated by Serf, a University of Delaware course system available by 1997, and by CourseInfo 1.5, built in 1996 by college students who later merged their company into Blackboard.
LessonCourse websites with logins and roles already existed in the mid-1990s, so a broad patent on them did not stand.
Credit
Desire2Learn (now D2L); US Court of Appeals for the Federal Circuit
2019
Six years of MOOC data
Massive open online courses (MOOCs), free courses open to anyone online, drew worldwide attention in 2012, the year Harvard and MIT founded edX. Justin Reich and José A. Ruipérez-Valiente analyzed 12.67 million registrations by 5.63 million learners in HarvardX and MITx courses up to May 2018. They found that 52% of registrants never opened the course material, that the share of learners who came back in their second year fell from 38% for the first cohort to 7% for the 2016–17 cohort, and that completion rates had not improved in six years.
LessonOpen enrollment alone did not bring in new learners; nearly all the growth came from the world's most affluent countries, and by October 2018 every major MOOC provider had turned to paid online master's degrees.
Credit
Justin Reich and José A. Ruipérez-Valiente, MIT, in Science
2024
An online program manager fails
2U, founded in 2008, helps colleges create and run online degree programs in return for a share of their program revenue, a business known as online program management. It bought edX for $800 million in 2021 and filed for Chapter 11 bankruptcy on 25 July 2024, under a deal with lenders and bondholders to cut its debt to about $459 million.
LessonColleges that hand their online programs to a contractor depend on that contractor's finances; the US Department of Education said it was "concerned" about the impact on students if 2U or other online program managers failed.
Credit
2U
Watch for
An AI agent inside Canvas
Instructure says its IgniteAI Agent for the Canvas LMS can build course modules and pages, generate grading rubrics, review discussion boards, reorganize modules and adjust due dates, drawing on more than 500 Canvas programming interfaces. It was free to US Canvas customers through 30 June 2026, and launch coverage reports no measured results, describing it as in an "early adopter phase".
Achieved whenAn outside party measures how often the agent completes instructors' requests correctly in real courses, and how much time it saves, and publishes the results.
Claimed by
Instructure, March 2026
Open
01
Course round-trip test
A public test of what survives when a course moves between learning systems. Build one reference course that uses every common item type (pages, files, quizzes with each question type, rubrics, discussions, links to outside tools, embedded video, image descriptions for blind readers), export it from each LMS in the Common Cartridge format, the shared course-package standard, import it into every other LMS, and record item by item what arrived intact, changed or vanished. Use only made-up course content and test accounts.
The briefCredit · unclaimed02
Agents on a real LMS
A rerunnable test of AI agents doing routine instructor work in a real LMS: extending a deadline for three named students, copying last term's course and shifting every date, rebuilding a gradebook category, posting an announcement to one section. Run it on locally hosted Canvas and Moodle filled with made-up students, and score each agent on whether the course ends in the correct state and on harmful side effects, such as a message sent to the wrong student or a grade changed by mistake.
The briefCredit · unclaimed
Open 01 · the brief
Course round-trip test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of what survives when a course moves between learning systems. Build one reference course that uses every common item type (pages, files, quizzes with each question type, rubrics, discussions, links to outside tools, embedded video, image descriptions for blind readers), export it from each LMS in the Common Cartridge format, the shared course-package standard, import it into every other LMS, and record item by item what arrived intact, changed or vanished. Use only made-up course content and test accounts.
Why it is still open
Vendors' help pages warn of losses but do not measure them. itslearning's import guide says some Moodle item types, such as glossaries, are "skipped during import", and lists lost formatting and missing images and tables among common problems; Populi warns that "every LMS includes (and excludes) different data" in its Common Cartridge exports. 1EdTech, the standards body behind the format, certifies products and keeps validators and test cartridges for its members, but publishes no comparison of what survives between systems. This research found no published side-by-side test across systems.
Built with what exists
The open-source code of Canvas, Moodle and Sakai, each run locally in Docker; trial or sandbox accounts for commercial systems where the vendor allows testing; the 1EdTech Common Cartridge specification; a Python script that unpacks each package's manifest file and compares items; screenshots for items that look different.
First step
Build a 10-item course in a local Canvas, export it as a Common Cartridge file, import it into a local Moodle, and list every difference.
Done looks like
A reference course of at least 40 items moved across at least four systems in every direction (12 or more pairs), with a published table marking each item intact, changed or lost, checked by two people, and the test rerun after each system's next major release.
Open 02 · the brief
Agents on a real LMS
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A rerunnable test of AI agents doing routine instructor work in a real LMS: extending a deadline for three named students, copying last term's course and shifting every date, rebuilding a gradebook category, posting an announcement to one section. Run it on locally hosted Canvas and Moodle filled with made-up students, and score each agent on whether the course ends in the correct state and on harmful side effects, such as a message sent to the wrong student or a grade changed by mistake.
Why it is still open
The closest test found, TeachArena (Chen et al., Hong Kong University of Science and Technology and Alibaba's Qwen team, revised September 2026), runs its 137 course-workflow tasks in a "deterministic, synthetic Canvas-style LMS", not in a real one; the best model scored 0.704 out of 1 on them. Instructure's own agent has no published success rate.
Built with what exists
Canvas's open-source code and Moodle in Docker; their REST programming interfaces; the community Model Context Protocol servers for Canvas on GitHub, which connect AI models to Canvas; a browser agent such as Browser Use for screens that have no interface (see the Note "Deleting 80% to start over"); Python scripts that compare the course before and after each task.
First step
Install Canvas or Moodle locally, load 30 made-up students, write 10 tasks with their target end states, and run one agent through them.
Done looks like
100 tasks with hand-checked target end states, each run five times by at least three agent setups; success rates and harmful side-effect rates published per task type, with the task files, so anyone can rerun them.