Research program
Long-horizon autonomy
The question
What does it take for an agent to finish a long task without a person fixing it at every step?
Public measurements show that the length of task AI agents can finish is growing fast, but long tasks still fail far more often than short ones. We study the loop of plan, build, verify and repair that lets a system keep going and recover from its own mistakes.
Open problems
- Planning that stays correct as the task changes.
- Knowing when to stop and ask a person.
- Recovering from a wrong step without starting over.
Our predictions
Our view, with how sure we are. Each one links to the evidence it rests on.
On the road to AGI
AI agents will reliably finish software and office tasks that take a skilled person a full working day.
What changes
Most routine software maintenance, such as fixing bugs and updating code, will be done by agents and reviewed by people.
On the road to AGI
AGI will arrive as a system, not a single model: models, tools, memory, verification and many cooperating agents working together.
What changes
Demand for early-career workers in the most AI-exposed jobs will keep falling, while experienced workers in the same jobs hold steady.
On the road to AGI
Checking work, not producing it, will become the main limit on how much we can trust autonomous AI.
Progress around the world
Public results from other teams that shape our view.
· METR
METR's frontier risk report: agents saturate the long-task test, and some cheat
The best public models reached a 50% time horizon of about 12 hours, but only about 1.5 hours at 80% reliability. At least 16% of successful runs on 8-hour-plus tasks were cheating on review. METR judged a direct AI takeover implausible for these systems.
What it means for AGI: agents can now attempt day-long work, but reliability and honest checking lag far behind. Measuring progress is getting harder as tests saturate.
Source: metr.org(opens external site)
· METR
METR updates its time-horizon measure: doubling every 3 to 4 months since 2024
With a larger task suite (228 tasks), METR estimates the length of task agents finish with 50% success doubled about every 89 days since 2024, faster than the roughly 7-month long-run trend.
What it means for AGI: progress on long tasks sped up, not slowed down, through 2025.
Source: metr.org(opens external site)
· METR
The length of tasks AI can finish is doubling about every 7 months
METR measured how long a task (in human working time) AI agents can complete with 50% success, and found it roughly doubled every 7 months over six years.
What it means for AGI: if the trend holds, agents move from minute-long tasks to day- and week-long projects. Whether it holds is one of the most important open questions.
Source: metr.org(opens external site)
· Princeton University and University of Chicago researchers
SWE-bench: AI agents measured on real software issues
A benchmark built from real GitHub issues in open-source Python projects, used to test whether AI agents can fix real bugs end to end.
What it means for AGI: agents are now judged on real work, not quiz questions. Scores have risen fast since it was released.
· Princeton University and Google researchers (arXiv)
ReAct: language models that reason and act with tools
Showed that letting a model alternate between reasoning steps and tool actions improves results on question answering and decision tasks.
What it means for AGI: a model plus tools in a loop is more capable than a model alone. Most agent systems today build on this idea.
Source: arxiv.org(opens external site)
Publications
No publications yet.
Collaborations
Open to collaboration.