The Development of AI Agents
Software systems that plan and carry out multi-step tasks on a person's behalf have moved from research demonstrations into workplace trials. The open questions are reliability, accountability and who bears the cost of mistakes.
Last meaningful change Oct 14, 2026, 9:00 AM · 5 milestones · 4 updates · 6 articles
Where the story stands
Software systems that plan and carry out multi-step tasks on a person's behalf have moved from research demonstrations into workplace trials. The open questions are reliability, accountability and who bears the cost of mistakes.
Background
For most of the past decade, machine-learning systems answered questions. Beginning in 2022, conversational models made that ability widely available, and within a year developers were connecting the same models to tools: web browsers, spreadsheets, code editors and company databases. An 'agent' in this sense is a model that is given a goal, allowed to take actions, observe the results and try again. The technical shift is less about intelligence than about autonomy and the error-handling that autonomy demands.
Timeline of developments
Documented milestones are dated and sourced. Items marked illustrative belong to the demonstration scenario.
· Origin
Conversational assistants reach the public
A free conversational model is released to the public and reaches tens of millions of users within weeks, making large language models a mainstream technology.
· Milestone
Models begin calling tools
A new generation of models is released alongside mechanisms that let them call external tools and services, the first step from answering to acting.
· Turning point
Computer-use demonstrations
A major developer releases a public beta in which a model can look at a screen, move a cursor and type, operating ordinary software the way a person would.
· Milestone
Browsing agents enter research preview
Recent meaningful changes
Each entry says what actually changed, not a rewritten summary.
- New developmentMajor
Employers describe agents moving from pilots to routine work
Illustrative: Several large employers told a workforce conference that software agents now handle first-pass invoice reconciliation and scheduling, with human review of exceptions. Previously these were described only as trials.
Read: As AI Agents Take On Office Work, the Question Shifts From Whether to How - Updated forecast
Productivity estimates revised downward
Illustrative: An economics group lowered its estimate of near-term productivity gains from agent adoption, citing the cost of reviewing agent output. Earlier coverage repeated the group's higher figure.
Read: Productivity Figures Hint at an AI Effect. Economists Urge Caution. - Contradictory evidence
Benchmark gains do not transfer to long tasks
Illustrative: A new evaluation found that agents scoring highly on short benchmarks completed under half of multi-hour tasks without intervention, complicating claims that reliability has been solved.
Coverage
As AI Agents Take On Office Work, the Question Shifts From Whether to How
Large employers describe moving software agents from pilot projects into routine back-office tasks. The harder problems are oversight, accountability and what happens to the people who used to do the work.
2 min read
Forecasts made about this story, revisited as evidence arrives.
- Open
Illustrative
Entry-level clerical job postings in early-adopting sectors will decline by a fifth within two years of agent deployment.
Illustrative labor-economics estimate · made Sep 1, 2025 · horizon 2027-09-01
- Open
Illustrative
By the end of 2026, agents will complete the majority of multi-hour clerical tasks without human intervention.
Illustrative vendor forecast · made Jun 1, 2025 · horizon 2026-12-31
Outcome: Illustrative: an October 2026 evaluation in this scenario found completion below half on multi-hour tasks.