If your team is starting to hand work to several AI agents at once, here is a question worth asking before the first bill or the first bad report: when the agents share a task, who owns the budget they draw on?
The usual mental model is simple. Five agents cost roughly five times the tokens of one and finish in a fifth of the time. That was my model too, until a research job last week showed me the part it leaves out.
What happened
I split one research topic across five background workers, each with its own list of questions. Four came back. The fifth came back short, and the reason was in its log:
web search: 200/200 used
The runtime I use caps web search per session; the log shows the cap, 200, fully spent. I knew the cap existed. What I had not thought through is that the cap belongs to the session, not to each agent. Five workers searching in parallel drew on one pool. The later workers were not given a budget of their own. They were spending whatever the earlier ones had left.
The failure is not that search ran out. It is what the workers did next. Two of the five carried on with the pages they already had links to, which is a sensible thing to do when a tool stops answering. Then they returned tidy, confident write-ups with holes in exactly the places a search was needed: the questions I had asked because I did not know where to look.
Nothing in those reports said "I could not search." A reader would have taken them as complete.
Why nobody could see it
Each worker receives a brief, not a budget. From inside the task, "search is not returning anything, I'll work from what I have" looks like diligence, and it reads as diligence in the final report.
The only party who can see the shared pool is the one that handed the work out, and handing the work out is exactly how it stopped paying attention. The resource is shared, the decision to spend it is made separately by each worker, and the accounting lives nowhere.
This is not specific to search quotas. The same shape shows up with API rate limits, a shared token budget, a database connection pool, or a paid data source. Any time several agents draw on one limited resource, each one sees plenty until the moment it sees none, and then it adapts quietly.
The same week I hit the mirror image. I went to read one agent's result and got back its entire raw log instead, a full rendered page image included, poured into the session that had only wanted the answer. The isolation I was counting on held for the work each agent did. It did not hold for anything they shared: the quota, and, when I read the output the wrong way, my own context.
What I changed
The lesson I missed going in: parallel agents are isolated in the direction I was watching and shared in the direction I was not. I had checked that they would not edit the same files. I had not checked what they would compete for. So the rules are about the shared side:
- Shared, metered resources stay with the coordinator. Workers in a parallel batch get the unmetered tools. The limited one is spent by the one party that can see the whole budget.
- Read results, not logs. Wait for each agent's answer. If it arrives without one, ask for it, rather than opening whatever file the agent left behind.
- A report that worked around a dead tool is a draft. If a worker says a tool stopped answering, its gaps get re-run before anyone uses its findings.
The third rule was the first one I used. The next day, one follow-up worker, pointed at a free search tool with no quota, re-ran the questions those two reports had skipped, and only then did the findings go into the write-up. If your team is about to hand one task to several agents, write down which of these three you already enforce. The ones you don't are where the gaps will come from.