Asymmetric Agents

Terence Eden, in a blog post titled 'Asymmetric Agents,' warns that autonomous AI agents will be vulnerable to manipulation by more powerful agents, citing the disparity between a low-complexity agent and Amazon's vast data centers. Eden argues that AI agents, like humans, will have biases that can be exploited, potentially by advertisers or malicious actors, undermining user control in a future of delegated tasks.

One of the promises of the AI-filled future is that we'll all have highly capable autonomous ~~servants~~ "agents" to work for us. Even back in the 1990s https://shkspr.mobi/blog/2026/02/how-close-are-we-to-a-vision-for-2010/ , it was a common trope in future-gazing to insist that a person's personal ~~slave~~ agent would negotiate on their behalf. Ask your ~~meek digital pal~~ agent to find you a restuarant in town, tonight, for a hot date, catering to your dietary preferences, and within your budget. The little ~~butler~~ agent scurries off and, without your intervention, finds the perfect place, negotiates with the restaurant's ~~house boy~~ agent and books you a slap up meal. Marvellous You can tell your ~~manservant~~ agent to buy you clothes, reserve a flight, do the grocery shopping, or buy your wife flowers without having to lift a finger. But, of course, there's a fly in the ointment. Not all ~~valets~~ agents are created equally. You're sending your Rasperry Pi powered ~~peon~~ agent against the full might of Amazons bazillion datacentres. Can a low-complexity ~~serf~~ agent be bullied, outwitted, and manipulated by a more capable ~~henchman~~ agent? It seems obvious to me that the answer is yes. Humans have a set of cognitive biases which render us susceptible to advertising tricks. AI may have different biases https://cacm.acm.org/research/rolling-in-the-deep-of-cognitive-and-ai-biases/ but will still be vulnerable to hypnosis https://shkspr.mobi/blog/2017/03/how-to-hypnotise-an-artificial-intelligence/ and hacking. Perhaps having my ~~secretary~~ agent negotiate with your ~~PA~~ agent to find a suitable meeting time for our Q2 KPI Deep Dive doesn't carry this sort of risk. But I'm sure that ~~scumbags~~ advertisers everywhere will be salivating at the thought of waylaying my innocent ~~footman~~ agent on its way to market. A future where we delegate responsibility to ~~stochastic imbeciles~~ agents is one where we accept that we have no control over their susceptibility to more powerful and malevolent ~~thugs~~ agents.