Blend's Agent Saves 4.5 Hours a Loan. The Number Alex Wang Asks For Is How Often Someone Stepped In.

Alex Wang's third question for long-running agents is how often people intervene. Blend reports hours and dollars. Here is the ledger that counts it.

Falk Gottlob · · 7 min read

AI Agents Falkster cover on green: a small navy robot walking along a curving cream track, and three large navy hands reaching down to touch the same stretch of track ahead of it, circled by an orange ring.

Long-running agents get measured in hours saved, and Alex Wang's November 12 edition asks for a different count. This is an Agent Drop: the ledger that produces the number in his third question.

The short version

  • Alex Wang's November 12, 2026 edition of Learn AI Together defines long-running agents as agents that keep making progress over time, and gives three patterns: event-driven execution, state kept outside the context window, and never trusting the agent's own "Done!"
  • His business case is Blend's Autopilot pre-underwriting agent, as he reports Blend's August release: more than 50,000 live loans since March, cycles 2 to 4 days shorter, 4.5 hours of fulfillment tasks automated per loan, and about $600 less cost per funded loan.
  • His third question for any company is how it will know the agent works, including how often people need to intervene. Blend's figures are days, hours, and dollars, and none is that count.
  • The drop is a step-in ledger: one row per human touch, with stage, kind, minutes, seen before, and what it became. Repeats are the number.
  • In my own scan, 13 of 42 reports recorded the same step-in nobody was there to make, and one line in the prompt ended it.

What Wang wrote

The edition starts from a task that does not finish in fifteen minutes. An onboarding agent prepares documents, waits two days for a manager's approval, and has to continue. The session is gone by then and some of the facts may have changed.

His distinction is clean. Multi-agent systems are about how several agents work together. Long-running agents are about how an agent keeps making progress reliably over time. You can combine them, and they are different problems.

Three patterns do the work. Know when to sleep: pause and wake on an approval, a schedule, or a tool callback instead of burning compute. Remember where you left off: persistent state and checkpoints outside the conversation. And do not trust the agent when it says it is done: automated tests, a separate evaluating agent, or a human on the critical steps. He walks through Anthropic's progress files and task lists, OpenAI's Agents API with automatic context compaction, and Cloudflare's dormant-and-wake design as three takes on the same problems.

Then Blend, and then three questions for any company. Where does work repeatedly get stuck. What can the agent safely handle, and which decisions still need a human. And how will we know it is working: not just tasks completed, but time saved, errors reduced, and how often people need to intervene.

The number nobody reported

Blend's figures, as Wang relays them, are 2 to 4 days, 4.5 hours, and $600. They are good numbers and they are all on the output side. Wang is careful about what they do not show: this is not one agent running every loan on its own for days. It is an agent doing real work inside a longer process that people are still part of.

Which is exactly why the count in his third question matters most. A loan that saved 4.5 hours and needed one planned approval is a different product from a loan that saved 4.5 hours and needed three unplanned rescues. The hours look the same. The second one does not scale, because the rescues grow with volume and the people doing them do not.

I made a related argument about migrations in Percent Migrated Is a Vanity Metric: the headline rate tells you what the system says, and the exception rate tells you whether it holds. Step-ins are the exception rate of a long-running agent.

The step-in ledger

One row every time a person touches a run. Six columns.

Run | Stage | Kind | Minutes | Seen before | Became

Kind has four values. A planned approval is one the design asked for, which is Wang's second question answered in advance. A correction fixes something the agent got wrong. An unblock answers a question or supplies access the agent could not get. A takeover is a person finishing the work.

Seen before is yes or no, and it is the column that pays for the ledger. A first-time step-in is discovery. You are learning where the workflow is soft. The same step-in a third time is a line missing from the harness.

Became has three values: a rule (a line in the prompt, a config value, a tool), a planned approval (you decided a human belongs there, on purpose), or nothing yet. Rows that stay at nothing are the backlog.

Two numbers come out of it. Unplanned step-ins per 100 runs, which should fall. And the repeat share, which should fall faster. If the first falls and the second does not, the team is firefighting, not learning.

One row, from this week

My creator scan runs every morning with nobody watching. It has written 42 reports since September 19, and each report ends with notes on what went wrong. Those notes are a step-in ledger that I had not been reading as one.

In 13 of the 42, the same thing happened. The browser connector found more than one Chrome on my account, first two, later three, then four, and asked which one to use. There was nobody to answer. The run fell back to a second browser every time and wrote a sentence about it. In the fallback, one writer's LinkedIn page kept hitting a login wall, so that source went unread.

13 of 42

scan reports recorded the same step-in nobody was there to make. One line in the prompt ended it.

Thirteen times. The same question. Seen before: yes, twelve times over. Became: nothing.

On November 12 I moved the schedule to another machine and wrote the browser's id into the prompt, with the instruction to select it without asking. On November 13 the run did, and it read the page the fallback could not.

That is what one row moving from repeat to rule looks like. It took one sentence. It took 13 reports to get written because nothing was counting repeats. The report format in Nielsen Says a Map Beats a Babysitter. Write the Map for the Next Run. had the facts in it. It did not have the tally.

The wake check

Wang ends on the part that is easy to skip. The longer a workflow runs, the more can change underneath it, and an agent that remembers exactly where it stopped may be resuming a plan that no longer makes sense.

That belongs in the ledger as its own kind of row. On every wake, before the next step, the agent re-reads the few facts its plan depends on and writes down whether any moved. My scan's version is the first instruction in its prompt: bring the repository up to date before anything else. When the wake check finds a change, that is a step-in the agent made on its own behalf, and it is the cheapest one in the table.

This week

Pick one agent that runs longer than a session. Go back through its last twenty runs and write a row for every time a person touched it. Sort by stage and mark the repeats. The top repeated row is this week's fix, and it is usually one line. This is the unglamorous half of Enterprise AI Agents: the agents that survive are the ones whose supervision cost was counted and driven down.

Related answer: How do you measure how often people have to intervene in a long-running AI agent?

Sources: Alex Wang, "We Know Multi-Agent Systems. But What Are Long-Running Agents?," Learn AI Together, November 12, 2026, read in full. The 50,000 loans, 2 to 4 days, 4.5 hours, and $600 are Blend's as he reports them from its August 2026 release, which I did not open; the Anthropic, OpenAI, and Cloudflare descriptions are his. The 42 reports, the 13, and the November 12 and 13 runs are from my own scan's records.

Share this post

Frequently asked

What is a long-running agent, in Alex Wang's definition?+

In his November 12, 2026 edition of Learn AI Together, Alex Wang separates long-running agents from multi-agent systems: multi-agent is how several agents work together, long-running is how an agent keeps making progress reliably over time, across sessions, context windows, and interruptions. He names three engineering patterns: event-driven execution (sleep until an approval, a trigger, or a callback arrives), persistent state and checkpointing outside the context window, and not trusting the agent's own claim that it is done.

What did Blend report?+

As Wang reports from Blend's August 2026 release, its Autopilot pre-underwriting agent had processed more than 50,000 live production loans since March, with loan cycles shortened by 2 to 4 days, an average of 4.5 hours of fulfillment tasks automated per loan, and an estimated $600 reduction in fulfillment costs per funded loan across the lender cohorts Blend analyzed. Wang adds that this is not evidence of one agent managing every loan autonomously for days.

Which number is missing?+

The one in Wang's third question: how often people need to intervene. Blend's reported figures are days, hours, and dollars. None is a count of step-ins per loan, and that count is what tells you whether the agent is getting cheaper to supervise.

What is a step-in ledger?+

One row every time a person touches a long-running run: the stage, the kind of step-in (a planned approval, a correction, an unblock, or a takeover), the minutes it took, whether it has been seen before, and what it became (a rule in the harness, a planned approval, or nothing). Report unplanned step-ins per 100 runs and the share that are repeats.

What is your own example?+

The creator scan behind this blog has written 42 reports since September 19, 2026. In 13 of them the browser connector found more than one Chrome and asked which to use, with nobody there to answer, and the run fell back to a second browser. On November 12 the browser's id went into the prompt. On November 13 the run selected it without asking and read a LinkedIn page the fallback had been hitting a login wall on.

Alex WangBlendlong-running agentsagent harnesshuman in the loopintervention ratescheduled agentsAgent Drops
Helpful?

Keep reading

The short answers

PM AI Agent Fleet, Mapped to the 7-Stage Operating SystemA live index of every AI agent for product managers, mapped to the 7 stages of the PM Operating System: Sense, Discover, Decide, Build, Ship, Measure, Amplify.Read the chapter

Part of

Enterprise AI Agents

About the author

Falk Gottlob

Falk Gottlob

Product Executive · Founder, Falkster.AI

Thirty years shipping product, from Microsoft Research and Adobe to Salesforce, where he grew Quip into what became Slack Canvas. Four startup exits, including a $6.5B healthcare merger and a company Microsoft bought. Four-time Chief Product Officer. Now founder of Falkster.AI, building Heidi, the company brain for AI-native businesses. This notebook is written from inside the build, not above it.

Sign in with LinkedIn to leave a comment.

Sign in with LinkedIn