Before you manage several AI agents, check that each one has a task worth giving it. Most failures start with the task, not the tooling. In one test on real freelance work, AI agents failed at 97% of real remote work tasks, and in another, the best system completed 26.2% of the assignments correctly.
Quick version
- A second agent does not fix a bad first task. Fix the task first.
- Warning signs: many steps that all must be right, messy inputs, heavy review, rare repetition, costly mistakes.
- Good fits: narrow, repeated, clear output, easy to check.
- Time the whole loop from request to finished work, not just the prompting.
- Add another agent only after the first one saves time on its own.
What the real work tests show
Two recent studies tried agents on real work instead of toy problems. Neither is your job, but both show the same pattern.
| Test | What it used | Best result reported |
|---|---|---|
| Remote Labor Index | 240 real freelance projects sourced from Upwork | Manus completed 2.5% of projects at a professional level |
| PYMNTS report on a research study | 1,490 real work assignments from more than 250 professionals across 55 industries | Codex on GPT-5.5 completed 26.2% of assignments correctly |
| PYMNTS report, hardest assignments | The hardest tier, which needs many steps over a long stretch of work | Pass rates across all systems averaged 2.6% |
The other models in the first test did no better. GPT-5 managed 1.7% and Claude Sonnet 4.5 managed 2.1%. On the hardest tier of the second, Claude Code on the Opus 4.7 model failed every hard assignment.
The projects were large. The average Remote Labor Index project took professionals 28.9 hours and cost $632.60. Our view: a small slice of your own work may do much better than a 28.9 hour project. Even so, the tests point at a clear rule. The longer the chain of steps, the less you can trust the result.
Easy work fares better, but not perfectly. On the easier tier of 59 assignments, top performers succeeded on roughly 30% of tasks, with Codex on GPT-5.5 reaching 42.4%. Even that best score leaves 57.6% unfinished, because 100 minus 42.4 is 57.6.
Signs you picked the wrong task
A task can look automatable and still cost you time. These are the common signs.
1. The job has many steps and every step must be right
This is the pattern in the hardest tier: assignments that require keeping track of many steps over a long stretch and getting every part right averaged a 2.6% pass rate. If one wrong step ruins the whole result, you will spend your time checking all of it.
2. The inputs are messy or live in your head
Seed & Society sells pre built AI agents and AI strategy consulting, so read its advice with that in mind. It says before you build or hire an AI agent, you need a clear process, clean inputs and a defined outcome, documented well enough that someone other than you could follow them. If you cannot hand the task to a stranger, an agent will struggle too.
Splintered Glass Solutions names the hidden cost. The context tax is the time people spend gathering, cleaning, translating or re-explaining information before useful work can begin. If you spend most of the effort feeding the agent, the agent is not saving you much.
3. Review and correction eat the savings
The same source says to track time from request to completed outcome, not just time spent prompting, and to include preparation, review, correction, approvals, handoffs and exception handling. A task that takes a minute to prompt and twenty to fix is a loss. Count the whole loop.
4. The task rarely repeats
Savings come from repetition. Seed & Society gives the math: a task that takes 30 minutes twice a week is 52 hours a year, and a task that takes 10 minutes daily is 60 hours a year. Setting up an agent costs effort once. A task you do a few times a year rarely pays that back. This is our judgment, not a sourced fact.
5. It means changing work that already runs
Maintenance is where agents can quietly break things. Alibaba's SUCCI benchmark, as summarized on DEV Community, found 75% of frontier AI models break previously working features when performing maintenance tasks. The same summary says maintenance accounts for roughly 60 to 80% of all software engineering effort. The lesson applies beyond code: an agent that edits something that works can leave you worse off.
6. A mistake is expensive
Seed & Society says client proposals, contract generation and payment processing, and anything with legal or financial implications, need a human review step. If you must read every line before it goes out, the task is a draft, not automation. Our longer take is in what to never automate with an AI agent.
7. The real work is judgment
MindStudio sells an AI agent platform, so weigh its framing accordingly. It argues that five constraints govern how much work an organization can capture: judgment, planning, coordination, evaluation and absorption. Notice that none of the five is typing speed. If the hard part of the task is deciding what is good, the agent does not remove it.
Where agents do earn their place
Some work fits. Seed & Society says the tasks worth automating are high volume, repeatable and time intensive, such as lead qualification, content publishing, meeting scheduling, proposal generation and email follow-up sequences. It also says low complexity tasks like scheduling, data entry and email triage are fast wins.
It offers one example, and it is the vendor's own account, so treat it as unverified. A business coach spent three hours a week publishing one blog article, and the time dropped from three hours to 15 minutes, a 92% reduction. The arithmetic checks out: three hours is 180 minutes, 180 minus 15 is 165, and 165 divided by 180 is about 0.92. Note what made it work. The steps were the same every week and the finished output was easy to see.
The Remote Labor Index also found a pocket of success. Agents showed success in audio production tasks like creating sound effects for retro games and mixing voice-overs with music. Narrow jobs with a clear finish line are where agents tend to land.
A test to run before you add a second agent
This is a general check, not a recipe for any one tool.
- Write the finished result in one sentence. If you cannot, the task is not defined yet.
- Time the whole loop by hand first. Use the same start and end points as the source above: request to completed outcome, including review and correction.
- Run the agent alone on that task several times. Compare the full loop time, not the prompt time.
- Write down what good and bad output look like. Splintered Glass suggests a small pattern library: approved inputs, desired output shape, source requirements, a review checklist, and examples of good and bad results.
- Plan to update the instructions. Seed & Society notes that agents do not learn in the background like a human employee, they follow the instructions you give them, so you must update the instructions to improve the output.
- Keep or drop it on evidence. The same Splintered Glass guide says to measure continued use, not just launch activity. If you stopped using it, it was not saving time.
Only then add the next agent.
Managing several agents without losing your evening
Agents can run in parallel. MindStudio notes that multiple instances of an agent can work in parallel, 24/7, on tasks you define. The catch is that you are the one defining, checking and absorbing the results. Each added agent adds review work for you.
There is a human limit too. One practitioner quoted by MindStudio says that with agents, instead of 8 to 10 normal productive hours, you might get 4 or 5 extremely intense hours before your brain is fully cooked. If you have one spare hour at night, that hour is the budget. Our view: two agents you can check in that hour beat five you cannot.
Cost is the other limit. In one version of the test reported by PYMNTS, researchers threw $630 worth of computing power and 763 million tokens at the problem and still succeeded just 2.9% of the time. More resources did not buy reliability. Set a ceiling before you start, as covered in AI agent spending limits.
If the reason you want agents is to free time for something of your own, the time you spend babysitting them counts against that goal. See how to ship side projects with a full-time job for how to protect that hour.
Keep the leash short
Because agents fail often on real work, the safe habit is a short leash: small tasks, clear checks, nothing touching your money without a human step. The Agent Leash is a one page checklist of 15 checks before your AI agent touches your money, sent the moment you sign up. Sunday Night Truths is free, one email every Sunday at 7 PM Central, in plain text, readable in two minutes. Get Sunday Night Truths.
