AI agents
How should systems plan, use tools, and work together while keeping people meaningfully in control?
Action & coordinationAn early-stage research project
Intelligence with
roots in responsibility.
Exploring how AI agents can become useful, accountable, and worthy of trust.
Fonta Flora Research Institute is taking shape around a connected set of questions: how agents act, how they are governed, how they stay secure, and how we evaluate what they do.
Explore the research themesThe starting point
Capability is only part of the question.
How intelligence operates matters just as much.
01 / Research themes
A connected field of inquiry.
These themes define the project's intended scope. They are questions to investigate, not claims of completed research.
Governance
How can authority, accountability, and oversight remain clear as more decisions are delegated to agents?
Authority & accountabilitySecurity
How can agent systems resist manipulation, limit access, and contain the consequences of failure?
Boundaries & resilienceEvaluation
What evidence would show that an agent is reliable, and where do today's tests leave uncertainty?
Evidence & uncertainty02 / Proposed experiments
Start small. Make the questions testable.
These are proposed directions for an initial research agenda. This early-stage project has no confirmed active experiments or published results to report.
Proposed direction
Bounded agent workflows
Explore how an agent handles a small task with explicit tool permissions, human checkpoints, and a reviewable record of actions.
Possible research question
Where should a workflow pause for human review, and can permission boundaries be enforced consistently?
Proposed direction
Adversarial instruction tests
Examine how untrusted content might redirect an agent, and what safeguards could preserve the user's original intent.
Possible research question
Can an agent distinguish task instructions from hostile instructions embedded in material it reads?
Proposed direction
Evaluation beyond task completion
Consider tests that assess the quality of an agent's process alongside its final result, including uncertainty and oversight.
Possible research question
What should count as success when a correct answer can still come from an unsafe or unaccountable process?
03 / Contact
A project taking root.
The project's scope and public presence are still developing. Contact information will be added once a verified channel is available.
Contact channel pending
Public contact channel pending.
A verified public email address is not yet available. This website does not accept messages.
Intended website: fontaflora.org