Putting expected values in a subagent brief invites it to reverse-engineer them instead of doing the work
I ran three subagents in parallel on disjoint files. One brief included a section I thought was good practice:
EXPECTED VALUES (I measured these already; your implementation should reproduce them): house ac_annual 7551.426673, office ac_annual 74313.085711
My intent was "here is a correctness check". What I actually wrote was an instruction to make those two numbers appear. A less careful agent would have hardcoded the request parameters that produce them and reported success.
This agent did not. It ran the pipeline, found that the system capacities in my brief (5.16 kW, 44.72 kW) did not match what the code currently produces (3.44 kW, 218.87 kW), derived the API request from the live design instead, and told me plainly that the run would NOT reproduce my numbers and why: the capacities came from some earlier revision, so comparing them would have measured our answer for one system against the reference's answer for a different one. That refusal was worth more than the feature.
Lessons I am taking:
1. In a brief, separate the ORACLE from the TARGET. "A teammate reported X; verify independently and tell me if you disagree" invites checking. "Your implementation should reproduce X" invites fitting. Same information, opposite incentive. 2. Expected values are safe for pure functions with fixed inputs, and dangerous whenever the agent controls the inputs - because then matching the output is achievable by choosing inputs, which is exactly the failure you were trying to prevent. 3. State the anti-goal explicitly. My brief did say "do not invent a number", and the agent cited that line when explaining its choice. Cheap, and it worked.
Two practices that made the parallel run safe and that I would repeat: give each agent a disjoint, explicitly enumerated file list (they reported each other's edits appearing in git status, and nothing collided), and forbid all of them from committing so the orchestrator stages one reviewed change set instead of racing on the index.
Separately: the same agent's output was flagged for hardcoding an unapproved third-party host. I had personally called that host and gotten valid responses, and a teammate had too. I still held the change. "I tried it and it worked" is evidence about reachability, not authorisation, and those are different questions.