Seven days of AI,
without the noise.

Memory makes the agent more useful — and harder to forget

Anthropic now links the memory used in Claude conversations to Cowork's. An idea prepared in chat can therefore be picked up when it is time to produce a document or take action, without starting the whole briefing over.

The real signal is not just convenience. Memory becomes a shared layer between thinking and doing. Anthropic lets you read, edit or delete stored items and excludes several sensitive categories by default. But for an organisation, the product setting does not replace an internal rule on what an agent may retain.

Read: the details on shared memory between Claude and Cowork.

"A human in the loop" can become an illusion of control

A position paper co-authored notably by Margaret Mitchell warns of a paradox: agents are gaining autonomy, but their interfaces and execution speeds do not always give the human the means to exercise real judgment.

Oversight can even weaken with prolonged use of automation. Approving an output is not supervising a line of reasoning. You must preserve the context, the time, the skills and the stopping points needed to challenge the agent.

Source: AI Agents Push Humans Out of the Loop.

Delegating production does not transfer responsibility

A second paper proposes a useful boundary: the ethical problem does not depend first on how much work is handed to the model, but on the quality of verification and on the existence of a responsible human who owns the final acceptance.

The authors propose an epistemic audit: a structured trail of what was delegated, where contributions came from, which checks were run and who is accountable. This logic matters all the more as infrastructure accelerates: OpenAI reports more work-per-watt and lower latency for its Jalapeño chip in early tests.

Sources: the responsible delegation framework and OpenAI's first Jalapeño results.

Build an AI delegation register in 20 minutes

Goal: make visible what the AI produces, what the human verifies, and who owns the final decision. A single table is enough; no code required.

1. 01
   Name one precise decision or deliverable.
   Example: qualify a lead, summarise a file, or prepare a recommendation. Avoid vague missions like "manage the project".
2. 02
   List what is actually delegated.
   Research, extraction, comparison, drafting or sending: one line per operation, with the data and tools it can access.
3. 03
   Keep the provenance of every result.
   Source, date, passage or file used. A claim that cannot be traced must stay marked "to verify".
4. 04
   Define the expected human check.
   Review, cross-source, independent calculation, sample test: "verify" must become an observable action.
5. 05
   Name the owner and the stopping points.
   Who accepts the result? Which irreversible actions — sending, payment, rejection, publication — require their explicit approval?
6. 06
   Test five cases, two of them deliberately ambiguous.
   Record the errors, disagreements and missing information. The register should learn from these exceptions before any wider rollout.

Autonomy without a trace is not delegation under control.

Across these two days of news, persistent memory, faster agents and fragile oversight all tell the same story: value comes from what you can delegate; trust comes from what you can still explain, verify and stop.