A few months ago, a single AI agent triggered nearly three hundred billion dollars in enterprise software sell-offs. The question worth exploring is whether that reaction was justified. So four of the most-discussed outcome agent tools — Cowork, Lindy, Sauna, and Opal — were run through a simple three-question framework. The best of the four scored roughly one out of three.
That result surprised me less than the reason for it.
Why code agents shipped first
The field has spent two years asking why AI works so reliably for software development and so unevenly for everything else. The answer, once you see it, is structural: code is verifiable. A compiler tells you whether it worked. A test suite tells you whether it broke something. CI — the automated pipeline that checks code changes — runs without anyone watching. The environment provides feedback automatically.
Knowledge work has no equivalent. When an outcome agent drafts a strategy memo, routes an email, or builds a research brief, nothing in the environment tells it whether it got the right answer. You are the test suite. That asymmetry — automatic verification for code, human judgment only for knowledge work — is why code agents work reliably and why knowledge work agents are still fragile. And it is not a maturity gap that will close next quarter.
This reframes the evaluation of any outcome agent tool.
The three questions worth asking
If the verification signal is missing from the environment, we need to look for structural properties in the tool itself that compensate for it.
Does the tool remember, or does every session start from zero? Genuine memory means the agent carries context from your last ten interactions, not just the last ten seconds. Most tools call this a memory feature. What’s needed is a memory architecture — a deliberate separation of permanent preferences, project-specific facts, and session history, each stored and retrieved differently. When memory is a feature the vendor controls, they own the compounding. When it is an architecture you control, you own the portability.
Does the agent produce something you can see, edit, and build on? The easier an interface is to use, the harder it typically is to debug. An agent that writes to a place you cannot inspect is not producing outcomes — it is producing opacity. Credits burn on failed tasks that quietly delete their own outputs. Inspectable surfaces are not a nice-to-have; they are what allows you to catch silent failures.
Does the tenth task go better than the first? Tools are useful; systems compound. If each run is roughly equivalent to the last, the agent is a faster search engine, not a collaborator. The tools that come closest to genuine compounding are the ones building loop structures — run, read what ran last time, write results back, repeat.
What the four tools actually scored
Cowork, which triggered the sell-off, scored roughly one out of three. Memory is treated as courtesy rather than architecture. Artifacts are fully inspectable — a clear yes. Compounding context is not there yet: it runs one task at a time and stops working when your laptop sleeps.
Lindy is the fastest path to feeling like something actually happened today, especially on email and scheduling. But the delegation model — describe the outcome, not the process — trades visibility for ease, and that trade surfaces exactly when something goes quietly wrong. One user consumed over two thousand credits on a task that failed and deleted its own outputs with no warning.
Sauna, which pivoted from a previous product after raising thirty million dollars, has the most architecturally sound foundation: memory as a substrate rather than a feature, recipe caching, proactive background jobs. It is early and unproven, but it is closest to how these tools need to be built.
Google Opal is free and uses Google Sheets as its memory layer — which sounds crude until you realize that memory you can open, edit, and inspect in a spreadsheet is more transparent than anything most dedicated agents offer. The tradeoff is platform risk and Google’s well-documented track record with new products.
What is actually durable here
The market has priced in a future where outcome agents replace significant portions of knowledge work. That future may well arrive. But it will be built on tools that solve the verification problem structurally, not tools that paper over it with cleaner interfaces.
The three properties that matter — memory you own, surfaces you can inspect, context that compounds — are stable regardless of which tools ship or pivot next. These are what to look for in any outcome agent, and what most current tools, even the best-reviewed ones, are still learning to provide. When one of them finally scores three out of three, that will be the real headline.
Text summarized and optimized using Anthropic’s models and reviewed by a human.