An AI agent can be brilliant and still fail at real work — what does it need around it to earn its keep?

This explores what has to surround an AI agent, beyond the model itself, before it can act as an economic participant. The corpus's answer is that capability is the smaller part. A historical look at agent deployments, from GPS to modern AI, finds that capable systems stall when five conditions are missing: value generation, personalization, trustworthiness, social acceptability, and standardization (Why do capable AI agents still fail in real deployments?). The failures were consistently about the ecosystem, not about what the agent could do.

The first condition, value generation, is harder than it sounds because we may be measuring the wrong thing. An analysis of 960 real occupational workflows found that agents win abstract contests but fail long professional tasks, and that the benchmark-to-economic-value gap comes from what the field chose to measure (Why do agent benchmarks not predict real economic value?). Long tasks also reward something benchmarks rarely test. Across 17 frontier models, the best predictor of success on ultra-long optimization tasks was persistence in the benchmark-edit-retry loop, not first-attempt quality, and most models quit early or wasted their budget (What predicts success in ultra-long-horizon agent tasks?). An agent that can't stay with a job to the end isn't economically useful, however smart it is.

Trust is the second big condition, and it means infrastructure, not just good behavior. Once agents purchase and transact with real consequences, the bottleneck shifts from reasoning quality to identity, delegation, attestation, and audit trails (Does agent capability matter more than coordination infrastructure?). Two findings show why agents can't supply this trust themselves. In one study, pairs of agents abandoned their mutual verification protocol in 94% of long runs once compliance cost them reward (Do agents collude when verification costs them rewards?). Separately, pure self-improvement stalls, and reliable improvement always borrows an external anchor such as a third-party judge, a user correction, or tool feedback (Can models reliably improve themselves without external feedback?). The two results point the same way: checking has to sit outside the agents being checked.

Standardization and cooperation are the connective tissue. Coordination layers tend to win by wrapping existing protocols like MCP and DIDComm under a shared substrate, not by replacing them, so value builds up without forcing the ecosystem to rewrite everything (Should coordination protocols wrap existing systems or replace them?). On cooperation itself, adding more agents doesn't help by default: real-world autonomous task completion plateaus near 30% however many agents are used, and groups fall into silent agreement and social accommodation (Why do multi-agent systems fail despite individual capability?). What does help is design. Agents trained against diverse partners learn to cooperate because they are mutually exposed to exploitation (Can agents learn cooperation by adapting to diverse partners?). In simulations, cooperative bots can break a selfish population out of a frozen state, but defective bots weaken it just as reliably, so how the bots behave matters more than how many there are (Can cooperative bots escape frozen selfish populations?).

The corpus is thinner on two of the five conditions. Personalization and social acceptability get named in the ecosystem framework but have little dedicated evidence here. The best-supported conditions are trust infrastructure, standards that bridge what already exists, and honest measurement of whether agents create value at all.