What it is
Anthropic’s engineering blog, written by an Anthropic engineer about the company’s own agent tooling.
The note describes two ways long-running assistants fail: trying to do everything at once and leaving half-built work, or declaring the job done too early. The fix is structure around the assistant rather than a smarter model. A first session sets up the environment and a progress file. The work is broken into a list of small, testable items in a file the assistant may only mark done, never edit. Each session starts by reading the log and the list, picks the highest-priority open item, finishes it, checks it end to end and records what happened.
The example is a large web application built this way. The result claim is that the approach produced production-quality software; no numeric benchmark is given, and the author says it is untested outside full-stack web work.
What marketing teams should take from it
- The setup around an assistant, which software teams call a harness, is what lets it stop at the right point and pick up correctly the next day. Finding 03.3 borrows that discipline for marketing.
- The checklist the assistant cannot edit is the important idea: someone else decides what counts as done.
- A readable progress log is what makes the work reviewable by a person who was not there.
- It is a pattern to implement with an engineer, not a product to buy.
Where to start
Use it to explain to legal and IT why approved sources, written instructions, a record of what the assistant did and a named person for exceptions are the controls, rather than a blanket approval queue.
Apply the shape to one long marketing job, such as checking 200 product claims against their sources. A reviewer owns the checklist, a systems person from marketing ops or IT sets up the log, and the functional head can read it.
Keep in mind
Written for coding work and untested elsewhere. Nothing in it covers brand or legal review; that layer is yours to add.