[01/06]Adaptation
How systems change through use.
We study how an AI system can learn from experience without drifting away from the person it serves.
Learning without driftBuilding AI systems that adapt to how people work.
Rifty is an independent research lab studying agents, adaptive systems, and the new forms of software developing around them.
Latest work from the lab.
Harness engineering with Codex: audit one coding task
Harness engineering with Codex starts with one coding task: supply usable context, bound its actions, inspect its changes, and leave a clear handoff.
Read
Agentic AI and AI Agents: A Primer for Leaders: a course-fit guide
The six-hour beginner primer covers fundamentals and basic custom GPTs; a scorecard tests its fit against production-readiness criteria.
Read
Agentic AI vs generative AI vs physical AI: compare the action boundary
Predictive, generative, agentic, and physical AI differ by what they produce, what they can change, and the controls their consequences demand.
Read
Agentic AI in banking: decide how much authority to delegate
Agentic AI in banking needs a six-class authority test for access, approvals, audit evidence, verification, and recovery.
Read
Is Copilot agentic AI? A capability-by-capability Agency Test
Copilot spans prompt-led generation, agent builders, workflows, and autonomous code changes; the Agency Test shows which capabilities count as agentic.
Read
Copilot Studio multi-agent orchestration: a control-first decision guide
Copilot Studio multi-agent orchestration needs narrow roles, bounded tools, tested handoffs, traceable execution, and a defined stop path.
Read
Our thesis
Software will stop being something everyone uses the same way.
Today's software encodes a workflow and asks people to follow it. As agents become more capable, that relationship can change. Systems can learn from the people they serve, reshape themselves around real work, and improve without waiting for another product release. The model supplies general intelligence. The surrounding system gives that intelligence a place, a memory, and a way to change. Today, that system is usually called a harness. We think it will become a new form of software.
What we study
Our work focuses on the systems that allow agents to learn, act over time, and become particular to the people and organizations they serve.
- [02/06]Memory
What should persist.
We study what should survive across tasks and how past experience should shape future work.
Across tasks - [03/06]Agency
Work that continues over time.
We study how agents pursue longer-running work across tools, interruptions, and changing conditions.
Long-running work - [04/06]Evaluation
Whether a system improved.
We study how to distinguish useful improvement from a system that has merely changed.
Evidence of progress - [05/06]Direction
How people shape the system.
We study how intent, taste, and judgment can guide a system without being reduced to a settings page.
Human direction - [06/06]Software
What comes after fixed workflows.
We study what becomes possible when software is formed around a particular person or organization.
Particular systems
Working assumptions
The ideas we are testing
People spend too much time preparing work for AI: moving information into a chat, restating context, and translating a job into language the model can follow. Better systems should take on more of that burden.
Two companies may use the same model and still need very different memory, tools, permissions, and ways of making decisions.
A system that changes itself is not necessarily getting better. Improvement only means something relative to the person, organization, or purpose it serves.
Models will keep improving and being replaced. The accumulated understanding of how a company works should survive those changes.
Operating environments contain missing information, changing standards, delayed consequences, and decisions that do not have one correct answer.
Experimental systems
Our research runs inside working companies.
Rifty maintains a private agent system that operates across tools, companies, and work that continues over time. It gives us a place to study what happens when tasks are ambiguous, requirements change, and the effects of a decision may not appear until days later. The system is not publicly distributed. We publish the parts that can stand on their own: findings, methods, evaluations, guides, and experimental tools.
Real environments
Research questions come from operating work rather than constructed demonstrations.
Continuous development
Failures, corrections, and changing requirements become new experiments.
Selective release
Work that generalizes beyond the private system is separated, documented, and published.
What we make available
- [01]
Research articles
Findings and arguments from our work on adaptive agent systems.
- [02]
Technical reports
Closer examinations of system behaviour, experiments, and design choices.
- [03]
Guides
Methods for people building systems around their own work.
- [04]
Evaluations
Ways to test memory, adaptation, reliability, and long-running agency.
- [05]
Experimental tools
Small releases for exploring one idea without adopting an entire system.
- [06]
Open questions
Problems we do not yet have good answers to, published so others can investigate them too.