Why do big companies seem to get AI agents working at scale faster than small ones do?

This explores why bigger companies seem to get AI agents working at scale before smaller ones do. One caveat first: none of these notes compares firms by size, so the premise isn't tested here. What the corpus does show is where the advantage in scaling agents comes from, and much of it comes down to money spent while the agents run. In multi-agent systems, about 80% of the difference in performance comes from how many tokens the system is allowed to use, not from smarter coordination between agents How does test-time scaling work at the agent level?. Research agents work the same way. Giving them more search steps improves results along the same curve as giving them more reasoning tokens, with similar diminishing returns Do search steps follow the same scaling rules as reasoning tokens? How does test-time scaling work for individual research agents?. If agent quality is largely bought this way, a firm that can afford a bigger running budget gets better agents. That's a plain, unglamorous reason larger organizations might pull ahead.

The second factor is less obvious. Making agents do real work is harder than winning benchmarks. An analysis of 960 real occupational workflows found that agents do well on abstract contests but fail at long, multi-step professional tasks Why do agent benchmarks not predict real economic value?. Closing that gap takes things large firms tend to have already: documented processes, people who can build the support systems around a model, and room to absorb failed pilots. Governance points the same way. Risk to people grows with the amount of autonomy handed to an agent, so the safer approach is to set graded levels of autonomy with human checkpoints Does AI risk increase with the autonomy we give it?. Building that kind of oversight is organizational work, and bigger firms have more staff to do it.

The corpus also points against size being a lasting advantage. Small language models can handle most agent subtasks at 10 to 30 times lower cost, which makes a mixed design (small models by default, large ones only when needed) the economically sensible choice Can small language models handle most agent tasks?. A 35B-parameter model trained on real execution data competes with much larger models while sitting near the cheapest point on the cost-performance curve Does model efficiency matter more than peak capability for real work?. Data needs can be surprisingly small too: 78 carefully chosen demonstrations beat 10,000 generic examples for training agent behavior How does test-time scaling work for individual research agents?. And most recent progress in self-improving agents comes from the 'fast loop', meaning updates to prompts, memory and tools, which are cheap and reversible, rather than from retraining model weights Do self-improving agents really split into two distinct loops?. A small team can do that kind of tinkering.

One more angle comes from a different direction. Farrell and colleagues argue that large AI models resemble markets and bureaucracies more than they resemble minds. They are lossy summarizers that compress and reorganize what people have accumulated Are large AI models actually intelligent agents or social technologies?. Read this way, an agent amplifies the information structure an organization already has. Large firms may scale agents faster partly because they already run on the formalized, documented, bureaucratic knowledge these systems are good at processing. The resulting picture: big firms' current lead looks like an advantage of budget and organizational structure, not of model access. The cost trends in the corpus suggest that lead could narrow for smaller organizations willing to build small, specialized systems instead of buying scale.