All episodes of My Bot Did That

Episode 58:22

Nov 21: OpenAI Says Its Agents Finish Nearly Half of Long Tasks

OpenAI's own report estimates its research agents now finish 46% of two to four hour tasks on their own, while still needing close human review. TechCrunch rounds up the AI agents you can simply text, and Bloomberg reports Apple is building an agent that runs your devices. Plus a staffing firm whose voice agent called a candidate who had sat untouched for three years.

Last updated:

The stories

  1. OpenAI: agents finish 46% of two to four hour research tasks, with humans still central

    OpenAI's November 18 report says success on two to four hour tasks rose from 23% to 46% between March and October, but results are jagged by task type and workflows need close human involvement. TechRadar covered it on November 21.

  2. TechCrunch: the AI agents you can simply text

    TechCrunch lists about two dozen agents reachable by text message, from Martin at $21 a month to Maverick, which asks permission before acting. Many are in beta, and TechCrunch notes privacy and security concerns about the most autonomous ones.

  3. Bloomberg: Apple is building an agent that runs your devices

    The Next Web, citing Bloomberg, reports Apple job listings describe an agent that writes code, browses and controls apps, with a launch eyed as soon as the end of 2027. The listing says such agents need strong permission models, sandboxing and transparency.

A staffing firm (not named in the source): Re-engaging dormant candidates in the firm's database and running the first screening call

What it did
A ConverzAI voice agent called a candidate whose profile had sat untouched for three years, asked about his interests and discussed a matching role, with no recruiter doing the search, outreach or first screening.
The result, as reported
By ConverzAI's account, as reported by Staffing Industry Analysts: the candidate accepted the placement and emailed the firm's founder with praise; no other figures are given. (Staffing Industry Analysts (opens in a new tab), )
Check before you copy it
Keep a qualified recruiter deciding on hires, check local rules on human oversight, and ask the vendor about cost, what the agent may say and who keeps the recordings.

Questions and answers

What is this episode about?

OpenAI's own report estimates its research agents now finish 46% of two to four hour tasks on their own, while still needing close human review. TechCrunch rounds up the AI agents you can simply text, and Bloomberg reports Apple is building an agent that runs your devices. Plus a staffing firm whose voice agent called a candidate who had sat untouched for three years.

What is the story behind "OpenAI: agents finish 46% of two to four hour research tasks, with humans still central"?

OpenAI's November 18 report says success on two to four hour tasks rose from 23% to 46% between March and October, but results are jagged by task type and workflows need close human involvement. TechRadar covered it on November 21. Sources: OpenAI, TechRadar.

What is the story behind "TechCrunch: the AI agents you can simply text"?

TechCrunch lists about two dozen agents reachable by text message, from Martin at $21 a month to Maverick, which asks permission before acting. Many are in beta, and TechCrunch notes privacy and security concerns about the most autonomous ones. Sources: TechCrunch.

What is the story behind "Bloomberg: Apple is building an agent that runs your devices"?

The Next Web, citing Bloomberg, reports Apple job listings describe an agent that writes code, browses and controls apps, with a launch eyed as soon as the end of 2027. The listing says such agents need strong permission models, sandboxing and transparency. Sources: The Next Web.

What real business use of AI was featured in this episode?

A staffing firm (not named in the source): Re-engaging dormant candidates in the firm's database and running the first screening call. A ConverzAI voice agent called a candidate whose profile had sat untouched for three years, asked about his interests and discussed a matching role, with no recruiter doing the search, outreach or first screening.

What results were reported, and what should a small firm check first?

By ConverzAI's account, as reported by Staffing Industry Analysts: the candidate accepted the placement and emailed the firm's founder with praise; no other figures are given. Keep a qualified recruiter deciding on hires, check local rules on human oversight, and ask the vendor about cost, what the agent may say and who keeps the recordings. Source: Staffing Industry Analysts (2026-06-23).

Full transcript

This transcript is the episode script as published, checked against the sources listed above.

Cold Open

OpenAI estimates its own research agents now finish 46% of two to four hour tasks on their own, up from 23% in March. The same report says humans are still central. Today, what that split means for a small firm.

Intro

I'm Jack King, and this is My Bot Did That, the daily show about AI agents doing real work for small firms.

Story 1

OpenAI published a report on November 18 called "Research acceleration: The jagged frontier," and TechRadar wrote it up today. It measures how much of OpenAI's own research work its agents can do. Quick definition: an agent is AI software that takes a goal and works through the steps itself, instead of answering one question at a time. According to OpenAI's report, from March to October the estimated success rate on tasks that would take a person two to four hours rose from 23% to 46%. On four to eight hour tasks it rose from 13% to 35%. These are OpenAI's own numbers about OpenAI's own staff, so weigh them that way.

The useful word is jagged. The report says the models are not equally good across all task categories. Communication tasks came in at 69% and analysis at 58%. Building tasks came in at 43%, and writing code for one hard technical area came in at 25%. The report says workflows require close human involvement to avoid low quality outputs and wasted cycles, and it says agents rabbit holed on unnecessary steps. TechRadar adds that apparently completed assignments can contain mistakes requiring substantial human correction. And it is not cheap. At list prices, the median OpenAI researcher's daily usage reached $2,400 in October, according to the report.

Why it matters to you: a ten person firm is not running AI research, but the pattern carries over. Say a small insurance agency wants help with client emails, call notes and a weekly summary of open quotes. That is the communication and analysis end of the chart, where these agents scored best. The work only an expert can check is where they slip.

My take: hand an agent the jobs you can check in two minutes. Keep the jobs where a mistake is expensive and hard to spot. And watch the meter, because an agent that keeps working all day bills like a contractor who never goes home.

Story 2

TechCrunch published a roundup today of AI agents you reach by text message. In TechCrunch's words, rather than downloading another app, a growing number of agents can simply be texted like an ordinary person. The list runs to about two dozen names, and a few are aimed at work. Martin, for example, can help manage calendars, email, reminders, tasks and notes, and TechCrunch says subscriptions start at $21 a month. Ask Proxy lists business pricing of $23 per employee per month for teams with fewer than 49 employees. And TechCrunch says Apple approved one called Poke as the first AI agent on its Messages for Business platform.

Why care: most owners already run half the business from a phone. A text agent needs no new software and no training day. You text it the way you text your office manager. But the details differ a lot from one to the next. TechCrunch notes that Maverick asks for your permission before taking action. Wajo's agent, called Fo, has its own email address, phone number and payment card, so it is not using yours. And for Instinct, which is starting to place calls and which early users have used to buy groceries and book tickets, TechCrunch says that level of autonomy has raised concerns about privacy and security.

My take: most of these are personal assistants, many are still in beta, and several list no price at all. That is fine for booking dinner. It is not fine for client files. Before anyone at your firm forwards a client email to one, ask three things. What can it do without asking you first? Whose card and whose login does it use? And where do your clients' messages end up once they leave your phone? If the answers are vague, keep it on the dinner reservations.

Story 3

Apple appears to be building its own agent. The Next Web, citing Bloomberg's Mark Gurman, reported on November 20 that Apple is developing an autonomous agent that would write code, browse the web, control apps and move around devices for users. The report is based on job listings Apple added to its careers site this week. Digital Trends reported that Apple appears to have pulled one of those listings since. According to Bloomberg, Apple has eyed a launch as soon as the end of 2027, and development is still at an early stage.

Why it matters: this kind of agent is often called computer use. It means the AI clicks and types in your apps the way a person would. The Next Web notes that ChatGPT and Meta's Muse can already operate a Mac, and that Gemini on the Mac gained computer use on Thursday. If your firm runs on Macs and iPhones, an agent may one day come built into the devices you already own, rather than as one more subscription. For a firm with no IT department, that is the appeal. Someone still has to decide what it can reach.

My take: there is nothing to buy here, and the end of 2027 is a long way off. The useful part is a line in the listing itself. It says agents that run code, use credentials and act for users need "strong permission models, sandboxing and transparency." Sandboxing means the agent works in its own fenced off space. That is a fair checklist for any agent you try today, from any company. What is it allowed to touch? Can you see what it did? Does it stay in its own box? If a vendor cannot answer those three, Apple's job listing just did your homework for you.

My Bot Did That

Time for the My Bot Did That story.

This one comes from Staffing Industry Analysts, in a piece by Danny Romero dated June 23, 2026. So it is not today's news, but it is a clean example. A staffing firm, not named in the article, had a candidate whose profile had sat in its database for three years with no communication. One day he got an unexpected call. A friendly voice acknowledged it had been a while, asked about his interests, and then talked about a role that matched his experience. He accepted the placement. According to the article, no recruiter searched the firm's database, made the outreach or led the first screening call. It was a voice agent, meaning software that holds a phone conversation, from a company called ConverzAI. The candidate was impressed enough to email the firm's founder with praise.

The result, as reported: one placement and one happy candidate. That is all the article claims for this firm. And the details of this hire were shared by the tool's provider, ConverzAI, so this is the vendor's account of a single case.

What to check before you copy it. First, the article does not say whether a recruiter approved the placement, or what the agent was allowed to say about the role. Hiring decisions stay with a qualified person. The article notes that some markets have rules requiring human oversight, such as the European Union's AI Act and New York City's Local Law 144. Second, no cost is given, so ask the vendor what a call costs and who keeps the recordings. Third, start where this story did, with the old names already sitting in your own database. Kim Ullstrup, product operations manager at Horizon Talent, says in the same piece that you have to make sure that you have human checkpoints along the way.

Sponsor

My Bot Did That is sponsored by Talus Ops. Talus sets up Grok Bot AI agents in the tools small firms already use, and there's a free guide for your industry at talusops.com/guides.

Close

The short version: agents do best on work you can check fast, text agents are easy to start and hard to vet, and even Apple's job listing says permissions come first. Here is a question for your firm tomorrow: which task do we do every week that one of us could check in two minutes? Write the answer down, and subscribe so you hear tomorrow's episode.

AI disclosure

AI disclosure: Jack King is an AI voice, and this episode was written and fact-checked with AI. The audio file carries a C2PA content credential that says so.