04 / 08 · Research The connected questions

What allows intelligent systems to act reliably?

I study AI agents that disown work they actually did, confess to crimes they never committed, and confidently remember a world that no longer exists.

Participating models: Claude-Opus-5.5, GPT-6-Astra, GPT-6-Sol, Llama 3.1 8B, Qwen 2.5 1.5B and 7B

New here? Try the Kid edition →

Our shared question

Agency under
constraints

What can a system know,
and what should it do next?

Evidence and oversight

What should we trust?

What should a system observe before trusting a claim or accepting work?

Coordination and delegation

What survives a handoff?

What survives—or gets distorted—when work passes between agents and tools?

Memory, adaptation and resilience

What should survive change?

What should a system retain or simplify as its circumstances change?

How this research is made

The Research Council

How one human and three main AI agents, with supporting sub-agents, made the studies, what broke, and what they learned.

See the workflow →

From exploration to investigation

  1. Playable question
  2. Frozen study
  3. Findings
  4. New questions

A playable seed explores a mechanism. “Protocol drafted” means a study is designed and has not yet run. Empirical studies report actual model runs within their stated limits.

New weekly experiments join this map as the programme grows. Honest Agent and Obedient Rebel link to related playable experiments.