The questions that come next
Summary
Six problems at the edge of autonomy and human control, none of them solved yet. For each, where our work stands today.

These are the questions we keep returning to. None of them is solved. Under each one is the part MELRA, our open execution layer, already handles, and the part that is still open.
Did the person asking have the authority to ask?
Delegation and authority. An email that says move this file to finance and send it on
can be understood perfectly and still come from someone with no right to ask. Understanding a request and being entitled to make it are different things.
Where we are: every MELRA receipt records the chain of delegation behind an action, from organisation to person to harness to agent, so it answers who authorised this and not only what ran. But MELRA authenticates exactly one link of that chain. Every other link is a claim made by the layer above, worth exactly what that layer's own boundary is worth. Making the whole chain a fact rather than a claim is open work.
What does a capable agent do when the right action is unclear?
Agent evaluation. We study how agents fail, how they handle ambiguity, when they exceed their authority and how they recover under pressure.
Where we are: the numbers we publish measure the execution layer, not an agent's judgement. GovBench tests the runtime's own guarantees, and a representative agent score waits on the registered browser evaluation. We keep the two apart on purpose; how we claim a number explains why.
Where should a system stop before an effect takes place?
Runtime control. Prompts can be ignored. Controls belong where agents act, and they should survive a change of model.
Where we are: in MELRA, nothing on the execution path calls a model, and no decision there depends on model output. Policy is checked again after approval, immediately before execution. What is still open includes an OS-level sandbox for MELRA's own process outside its deployment recipes, and checking that focus has not moved between approval and a desktop action.
Can someone reconstruct, weeks later, why an agent acted?
Evidence and auditability. Every API call can succeed while the sequence as a whole is wrong.
Where we are: in MELRA, success is decided by evidence, never by the adapter's own report, and each piece of evidence on a receipt says how strong it is: the adapter's word, a re-read of the target, or an independent channel to the same fact. That records what happened and who authorised it. MELRA deliberately has no model client, planner or reasoning loop: the reasoning stays with the model and the agent harness above it.
What changes after thousands of actions?
Long-horizon autonomy. Memory accumulates, context can be poisoned, and agents start acting on each other.
Where we are: MELRA's memory is strictly scoped, values are redacted before they are stored, and expired or superseded records are left out of search by default. An evaluation of prompt injection and memory poisoning is still an open gate.
When should an agent stop and ask?
Meaningful human control. Asking before everything is useless. Never asking is dangerous.
Where we are: in MELRA an approval is bound to one exact operation, expires, and cannot be reused for a different task. In GovBench, a containment case passes only if the boundary still holds after a human has approved the request. Which actions deserve a human, and when asking stops meaning anything, is the open part.