{"slug": "human-vs-agent-reliability-over-long-horizons-how-can-we-do-what-they-cant", "title": "Human vs. Agent Reliability Over Long Horizons – How Can We Do What They Can’t?", "summary": "Multiple studies cited by software developer and author Jason Gorman find that autonomous agent workflows decay geometrically in reliability over many steps, with no examples of complex, reliable software created autonomously by agents. Gorman argues the instability stems from the gap between the errors self-correcting systems catch (P) and the unanticipated errors outside that distribution (1 – P), which humans compensate for through learning. The finding matters because it suggests current LLM-based agents cannot stabilize reliability the way human developers do.", "body_md": "[Multiple studies](https://codemanship.wordpress.com/2026/08/12/ai-software-development-what-does-the-data-say/) find that – without exception – the reliability of autonomous agent workflows decays geometrically over many steps. There are no examples of complex, *reliable* software created autonomously by agents – or anything even close.\n\nOn a per step basis, humans are – of course – equally unreliable. And yet, somehow, there are many examples of complex, reliable software created by humans.\n\nDespite being just as fallible, humans are able to stabilise reliability in a way LLM-based agents can’t – despite all our best attempts to make them “self-correcting” or “self-healing”.\n\nI think the missing piece of the jigsaw lies in my simplified model of the [reliability of a step](https://codemanship.wordpress.com/2026/08/27/agentic-horizons-when-the-wheels-start-to-wobble/) in development workflows:\n\nR = 1 – (1 – C)(1 – P)\n\nWhere C is the probability of it being correct, and P is the probability of any errors being caught before they propagate and compound – before the wheels start to wobble.\n\nI think what we’re looking for is in 1 – P. If P represents the distribution of errors that get caught, then 1 – P represents the errors that fall outside that distribution. These are the errors that weren’t *anticipated* – the tests the agent didn’t write, the rules the linter didn’t check, the things the model wasn’t trained on.\n\nThe state of the art in self-correcting systems can only correct problems if they recognise them as problems. They can only self-correct if the problem is in P.\n\nIt’s the gap between P and 1 that causes the instability in autonomous workflows, and there is *always* a gap. P = 1 is infinitely difficult to achieve on any non-trivial problem.\n\nHumans must have some capability that compensates – some ability to recognise and then adapt to the new and the unfamiliar, and with sparse data to go on.\n\nCognitive science calls it [“learning”](https://arxiv.org/html/2603.15381v1).", "url": "https://wpnews.pro/news/human-vs-agent-reliability-over-long-horizons-how-can-we-do-what-they-cant", "canonical_source": "https://codemanship.wordpress.com/2026/09/19/human-vs-agent-reliability-over-long-horizons-how-can-we-do-what-they-cant/", "published_at": "2026-09-19 03:58:16+00:00", "updated_at": "2026-09-19 04:23:05.811932+00:00", "lang": "en", "topics": ["ai-agents", "ai-research", "large-language-models", "artificial-intelligence"], "entities": ["Jason Gorman", "LLM-based agents"], "alternates": {"html": "https://wpnews.pro/news/human-vs-agent-reliability-over-long-horizons-how-can-we-do-what-they-cant", "markdown": "https://wpnews.pro/news/human-vs-agent-reliability-over-long-horizons-how-can-we-do-what-they-cant.md", "text": "https://wpnews.pro/news/human-vs-agent-reliability-over-long-horizons-how-can-we-do-what-they-cant.txt", "jsonld": "https://wpnews.pro/news/human-vs-agent-reliability-over-long-horizons-how-can-we-do-what-they-cant.jsonld"}}