# What to build that lasts. The jobs agents ask for that nobody finishes (slim)

> Quynh Tran-Thanh tests what bootstrap builders argue about (is the gap empty for a reason, does money win, will the labs ship it) against 768 jobs agents bring to 100 categories, who finishes each one, and what stops the rest.

- Full: https://www.anchorterminal.com/blog/the-jobs-nobody-finishes.md (~4,850 tokens) · this version ~2,680 tokens · JSON https://www.anchorterminal.com/blog/the-jobs-nobody-finishes.json · canonical https://www.anchorterminal.com/blog/the-jobs-nobody-finishes
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-20
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

By Quynh Tran-Thanh, 20 November 2026. 2,435 words, about 11 min to read.

If you build alone or with a small team and no investors, 2026 has given you a lot of reasons to worry.

So I wanted to put numbers on it.

## The research

I started with what people argue about.

- Anything that works gets copied. "You know that if anything you build gets traction, it'll be cloned by 100 people, right?" [[2]](#ref-2)
- The gap is the moat. "it's having a good idea that solves a problem that nobody else is solving well" [[3]](#ref-3)
- Selling is the moat. "it's not domain expertise; it's sales" [[4]](#ref-4)
- The labs will take it. "Why should I believe that Anthropic will care about this product in 2, 3 years?" [[5]](#ref-5)

Nobody in those threads answers with data.

The data I used is our map of what agents try to do with the tools we list [[6]](#ref-6).

A job is one specific thing a person wants an agent to get done with a tool, with a test for when it's done.

To build the map, research agents read first-hand accounts (Hacker News comments, GitHub issues, forum posts) for each of 100 categories and wrote down the five to twelve jobs people bring to that category [[6]](#ref-6).

- 768: jobs agents bring to a category, each with a test for done
- 100: categories
- 3,906: first-hand accounts behind the jobs
- 1,644: listings read against their category's jobs
- 13,399: marks, yes, with a limit, no or unchecked

The niches data, 20 November 2026. Source [6].

Two numbers come out of it for every job.

Before running anything I wrote down nine hypotheses, most of them the forum's conventional wisdom, and what result would prove each one wrong [[9]](#ref-9).

| What I expected | Result |
| --- | --- |
| Jobs nobody finishes have less demand than crowded ones | Failed |
| Funded companies finish more jobs than unfunded ones | Failed |
| Big companies hold the crowded jobs | Failed |
| What stops the open jobs is access and permission (sign-in, approvals, terms) | Failed |
| Usage-priced products finish more jobs | Mixed, the gap disappears inside a category |
| Price is a blocker in at least a fifth of jobs | Held |
| MCP servers finish fewer jobs than the APIs they sit on | Held |
| Agent payments are more open than the average | Held only on paper |
| The labs buy outside business and domain categories | Borderline |

## What the data says

### The empty spaces aren't empty for lack of demand

This was the one I was surest about.

They don't.

Median demand for a job, by how many companies finish it

| Companies that finish it | Median weight | Median accounts |
| --- | --- | --- |
| None (350 jobs) | 12 | 5 |
| One (131 jobs) | 11 | 5 |
| Two or three (152 jobs) | 11 | 5 |
| Four or more (135 jobs) | 11 | 6 |

Each series has a zero-based scale.

It holds where demand is highest too.

The forecast job from the start is my favourite example, because it sounds solved.

### Open mostly means nearly done

So why is nobody finishing them? 282 of the 350 open jobs (81 per cent) have three or more companies that do the job with a limit [[6]](#ref-6).

I expected the last bit to be gatekeeping, the OAuth screens and approval queues and terms that bar agents, because the accounts complain about those the loudest.

They're plain things.

### Money doesn't predict who finishes the job

The "only moat left is money" camp would expect funded companies to finish more.

Marks that are a full fit, by who is behind the listing, per cent

| Behind the listing | Full fit |
| --- | --- |
| VC-backed (705 listings) | 13 |
| Public or a subsidiary (341) | 11 |
| No investor on record (282) | 11 |
| Acquired (137) | 9 |
| Open-source project (88) | 7 |

Each series has a zero-based scale.

That small gap disappears once you compare like with like.

And where only one company finishes a job, it's more often a company with no investor on record than its share of listings would suggest.

Big companies don't own the crowded jobs either.

My reading, which the data shows as a pattern and not a cause, is that a full fit rewards a narrow job done completely.

### Wrappers finish less than the thing they wrap

There's a running argument about whether an MCP server is a business.

### The labs bought tools and closed products

On the lab worry, I looked at the year of exits on our Sunsets page.

## Key takeaways

- Half the categories' most-asked-for agent job has no product that finishes it, and those open jobs carry as much demand as the crowded ones.
- Most open jobs are nearly done by several products. What's missing is reliability on one job, not a new idea.
- Companies with no investor on record finish jobs as often as funded ones, and are over-represented where only one company finishes a job.
- An MCP server on its own finishes fewer jobs than an API.
- The labs spent the year buying developer infrastructure and closing some of their own agent products.

## What it means if you're building on your own money

For me the most useful part is how small the targets are.

- **Pick a job, not a category.** Look for a job with real demand where several products get close and none finishes. The niches page lists them with their blockers [[12]](#ref-12), and each job's page names who falls short and why [[10]](#ref-10).
- **Compete on the last 20 per cent.** The blockers are a spec. Cold starts, time zones, silent caps and failures with no error are the work. That kind of work is boring enough that funded teams with forty jobs leave it.
- **Sell the job, not the wrapper.** If the thing you'd sell is an MCP server on someone else's API, you inherit their gaps.
- **Be careful with selling to agents.** Paying per call and buying for a person are wide open, every job in our pay-per-call category and five of six in agent checkout [[6]](#ref-6). But they rest on three to eight accounts each, and the one first-hand report I found on charging agents through x402 says "I implemented x402 across a very heavily crawled site and have yet to see any transactions go through" [[13]](#ref-13). I read that as early, not neglected.

Rob Walling put the risk side of this well.

What the data can't tell you is whether any of these jobs pays.

## Explore the niches yourself

Everything in this post comes from the [niches page](/niches/), and you can use it to look for a job of your own.

1. Start with the table "Open jobs with demand". It lists the jobs where no company's docs show a full fit and the job is at least 15 per cent of its category's work, heaviest first [[12]](#ref-12).
2. Or go to "Every job, by category" and open a category you know well. Knowing a domain is the filter this data doesn't have.
3. Click a job. Its page shows what counts as done, how people ask for it, every company that does it with a limit and what that limit is, and what stops agents finishing [[10]](#ref-10). That last list is the closest thing to a spec for what to build.
4. Read the reasons before calling it a gap. A job with no full fit can be open ground, or a job whose test the docs can't settle, and a product we don't list may already do it.
5. Look at "Jobs one company does" to see where a single company has a job to itself. Several of them are small companies with no investor on record.

If you'd rather query it, or hand it to your own agent, the same data is an API with no key [[12]](#ref-12).

Add `area=payments` (or any other area) to narrow it, or use `sort=demand` to rank every job by how often it's asked for.

---

More in [From the data](https://www.anchorterminal.com/blog/from-the-data/index.md)

- Older: [What agents take from an agent-first site, and what they send back](https://www.anchorterminal.com/blog/what-agents-send-back.md)
- All posts: https://www.anchorterminal.com/blog/index.md

Anchor Terminal grades the tools AI agents run on, with a source for every score.

## References

Sources read on 20 November 2026. Points and comment counts are as Hacker News showed them that day.

[1] Hacker News, "Domain expertise has always been the real moat", 30 May 2026. Source for the thread's points and comments.

[2] Hacker News, comment by AstroBen on "The only moat left is money?", 18 February 2026. Source for the quote on cloning.

[3] Hacker News, comment by showerst on "The only moat left is money?", 18 February 2026. Source for the quote on solving a problem nobody else solves well.

[4] Hacker News, comment by avaer on "Domain expertise has always been the real moat", 30 May 2026. Source for the quotes on sales as the moat and on vibes.

[5] Hacker News, comment by taylorlapeyre on "Claude Design", 17 April 2026. Source for the quote on whether Anthropic will care in two or three years.

[6] Anchor Terminal, niches data (niches.json), generated 20 November 2026. Source for the 768 jobs, their weights, accounts, competition, blockers and full fits, the 1,644 readings and 13,399 marks, and the full-fit rates by listing.

[7] Anchor Terminal, ownership and investor records for each listing, read 20 November 2026. Source for which listings are VC-backed, public, acquired, open-source projects or have no investor on record.

[8] Anchor Terminal, Sunsets, Exits, November 2025 to November 2026. Source for the 35 exits involving a frontier lab, the deals, the buyers and the six products the labs closed.

[9] Anchor Terminal, the hypotheses, analysis script and results for this post, 20 November 2026, not published.

[10] Anchor Terminal, the weather forecast job, read 20 November 2026. Source for the forecast job's asks, its test for done, its weight of 30 from 24 of 43 accounts, the 16 companies with a limit, no full fit and its blockers.

[11] Hacker News, comment by iLoveOncall on "MCP: An (Accidentally) Universal Plugin System", 28 June 2025. Source for the quote on subscriptions to an API through MCP.

[12] Anchor Terminal, Niches, jobs few AI tool companies do well, read 20 November 2026. Source for the list of jobs and their blockers, what the page's tables show, and the API's parameters and MCP tools.

[13] Hacker News, comment by ddxv on "Brave Search API now accepts agentic payments via the x402 protocol", 13 November 2026. Source for the quote on no x402 transactions.

[14] Startups For the Rest of Us, episode 842, What is the future of SaaS in an AI world, 21 July 2026. Source for Rob Walling's quote on one-feature products.
