Published Briefings

Meeting preparation with Cowork: connecting evidence, analysis, and action

From email to analysis, slides, and calendars: what must stay connected when one assignment spans the work?

AI chat UI: when to add an editor or a visible plan

A conversational request and a precise edit solve different problems. Compare drafts, shared versions, editable plans, and recurring forms.

Agent security: trace external instructions through to execution

Recognizing an injection does not establish that transmission stopped. Verify that permitted work can still finish.

Enterprise AI agents: define the job and who takes it back

A successful ordinary request leaves exceptions to handle. Include the receiving person and remaining work in the adoption criteria.

Comparing AI agents: models, retrieval, and trial allocation

Did the new model help, or did another search? Separate explaining the gain from choosing the configuration to use.

Agent updates: distinguishing improvements from regressions

Answering a new question does not prove that an established escalation survived. What evidence should decide acceptance?

AI agents in 2025: what changed the work we can delegate?

More components became available. Choose the next task to delegate by whether its output advances human judgment.

Foundry multi-agent workflows: passing values with evidence

A well-formed value need not be a fact the next agent may trust.

AgentKit and agent toolkits: from quick prototypes to verified releases

Saving a workflow does not ensure that its policy source and interface reflect the same change.

Agent SDKs in business workflows: adaptive steps, fixed conditions

Allowing an AI to explore does not require letting it change revenue definitions or submission conditions.

From AI-written code to comparable experiment results

A good score alone cannot settle adoption. Connect data, conditions, and results, retaining experiments without improvement too.

AgentOps after a connection loss: deciding what can be retried

History alone cannot justify a retry. Distinguish committed changes from unknown outcomes so people can continue the work.

MCP and A2A: what must survive a change of provider

After replacing a search provider, can users still interpret its results under the same conditions?

Multi-agent research: dividing work that leads to one usable report

Faster collection can leave more checking to the user. Choose division from the dependencies needed to finish the report.

A2A handover: what must travel with an agent request

Connect delegation to request meaning, validation after edits, and the progress users need to see.

OpenAI and Bedrock: what work to delegate to an agent platform

Connecting tools and specialists still leaves work to finish. Compare what the platform executes with what the application must supply.

Agent evaluation: when a higher score is not enough to deploy

An improved average leaves deployment questions open. Read repeated failures, grading conditions, and the effort required from users.

Browser agents: using existing screens without losing the job

More screens mean more outcomes to verify. Compare the breadth of browser access with the effort needed to continue each job.

Evaluating AI agents: a successful tool call is not a finished job

A successful search may still leave the person’s question unanswered.