Scope Note
This article uses responsibility as an operational allocation concept: who executes, who reviews, who owns the outcome, and who handles exceptions. It does not claim that a normalized vector can replace legal accountability, managerial accountability, or organizational duty. Those can remain fully with the deploying organization even when the internal routing weights change.
1. The real problem
Most responsibility failures in agent systems do not come from having no logs. They come from having many contributors and no clear owner. One agent drafts, another verifies, a third executes, and a human approves only edge cases. After an error, everyone touched the case but nobody clearly owned the final judgment path.
A usable model therefore needs to answer four questions for every decision: who acted, who checked, who had authority to stop it, and who receives the next escalation if something goes wrong.
2. A practical responsibility vector
For a decision d, define an operational vector rho(d) = {rho_exec, rho_review, rho_owner, rho_exception} with an optional human reserve rho_human. The normalized budget satisfies rho_exec + rho_review + rho_owner + rho_exception + rho_human = 1.
The point of the budget is not philosophical neatness. It is to make missing functions visible. If rho_owner = 0, nobody owns closure. If rho_review = 0 on a risky workflow, nobody is explicitly checking quality. If rho_exception = 0, failures have no natural landing point.
3. Allocation patterns
Low-risk repetitive work
Use lightweight execution-heavy allocations. Most weight can sit on execution and ownership, with small review and human reserve. This keeps friction low while preserving a clear stop path.
Specialized multi-agent work
Use competence-weighted review and ownership. The agent with the deepest domain signal may deserve more review weight, but ownership should still remain stable enough that someone is accountable for the final merge.
High-ambiguity or high-impact work
Increase rho_human and rho_exception. The right response to ambiguity is not to spread tiny fractions across many agents. It is to reserve explicit review and override capacity.
4. What the conservation idea is useful for
The earlier version of this article described responsibility as a strict zero-sum transfer between agents and humans. That framing was too broad. What is truly conserved here is the internal routing budget, not legal accountability in the outside world.
The useful invariant is narrower: every required function in the decision path must be allocated somewhere. Within that normalized budget, raising one share lowers another. That is operationally helpful because it forces designers to admit tradeoffs instead of pretending every participant fully owns everything.
5. Avoiding bad allocations
Three patterns are consistently unhealthy. First, diffuse ownership: five agents each own a little, so no one owns closure. Second, review mirroring: the executor and reviewer are effectively the same capability, so review adds latency without real independence. Third, chronic concentration: one trusted agent ends up owning too many decisions and quietly becomes a single point of organizational failure.
A simple monitoring approach is to track both concentration and mismatch. Concentration can be measured with HHI or Gini on ownership weights by workflow family. Mismatch can be measured by checking whether the reviewer has materially different evidence sources, tools, or incentives from the executor.
6. Adaptive reallocation
A lightweight adaptive rule is often enough. Start from a baseline allocation template by workflow type. Then adjust weights from observed error discovery, reversal cost, and escalation frequency. If review keeps finding defects late, raise review share earlier in the flow. If human queues back up without reducing rework, lower human reserve for that workflow and move review deeper into the agent team.
This is a heuristic control loop, not a theorem. The objective is stable ownership and timely checks, not mathematical elegance.
7. Internal replay findings
In internal replay and synthetic workload tests, moderate human reservation performed better than either extreme. Keeping some explicit human reserve reduced silent drift and late reversals, while routing every case to humans created queueing without proportional quality gain. The most useful operating band depended on domain risk and evidence quality, but often sat in the 0.2-0.4 range for ambiguous business workflows.
The stronger claim that a universal optimum exists was not defensible. The correct conclusion is narrower: reserve human budget where reversibility is low, evidence is weak, or policy interpretation is unstable.
8. Deployment checklist
- Make the owner field mandatory for every completed decision
- Separate executor and reviewer capabilities on material workflows
- Keep an explicit exception handler rather than implicit fallback to ops chat
- Track concentration of ownership over time by workflow family
- Increase human reserve only when it reduces reversals or unsafe approvals
Conclusion
Responsibility allocation is a workflow design problem. The right goal is not to compress accountability into a single abstract number, but to ensure that execution, review, ownership, and exception handling are always assigned somewhere explicit. A normalized responsibility budget is useful precisely because it exposes tradeoffs and makes missing roles visible before failures do.