AI uses for records
Human-agent disagreement resolution data: recording a corrected plan
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Training data should capture a human's disagreement with an agent's plan as a linked set: the plan as proposed, the decision type (approve, edit, reject, answer instead, decline or dismiss), who decided and under what authority, the constraint the correction adds, the plan that actually ran, and the stored record that shows the final outcome. A rejection flag alone teaches nothing about why or whether the fix held.
Agent run
Human agent disagreement
Key takeaways
- Record the decision type exactly; an edit, a rejection with a reason, an explicit decline and a dismissed prompt teach different things.
- Store the authority behind every correction, because a disagreement from someone without decision rights is evidence, not a label.
- Turn each correction into a written constraint and check that it is still present after every checkpoint and handoff.
- Prove the outcome with the stored business record, not with the agent's closing message or a status field.
- Keep counterexamples: corrections that were wrong, and disagreements that should have been escalated instead of obeyed.
How should training data capture human disagreement with an agent's plan?
As one linked case, not a single label. The case starts with the state of the business records and the instruction the agent received, holds the plan it proposed, records exactly how a person responded and on whose authority, and ends with the stored record that shows what finally happened. Each part references the others by identifier so a reviewer can walk from the disagreement to the outcome.
The reason is that disagreement carries three separate lessons. The first is about intent: the person knew a requirement the agent missed. The second is about authority: the person was entitled to change the plan, or was not. The third is about durability: the correction either held through the rest of a long task or was lost at a checkpoint, a retry or a handoff to another agent or shift. A dataset that keeps only approve or reject cannot teach any of the three.
This guide sets out the record, the decision types worth separating, how to capture authority, where corrections get lost, an illustrative procurement case, and the counterexample every pack needs.
Why does an approve or reject flag lose the lesson?
Because the frameworks agents run on already distinguish more outcomes than two, and each one means something different for training. LangChain's human-in-the-loop middleware defines four decision types: approve, edit, reject and respond. A rejection can carry the person's reason. An edit replaces the agent's tool call before it runs, and the middleware then tells the model that the call in its own message is not the one that executed. A respond decision answers on the tool's behalf without running it.
The Model Context Protocol adds a distinction that matters for disagreement. When a server asks the user for input, the response is accept, decline or cancel, and the specification defines decline as an explicit refusal and cancel as dismissing the request without making a choice. Its tools specification also says a person should be able to deny any tool invocation. A dismissed prompt is not a disagreement, and treating it as one teaches the agent that a person objected when they simply closed a dialog.
So the first field in the record is the decision type, recorded with the vocabulary of the system that produced it and mapped to a shared set. The table below is the mapping the task pack uses.
| Decision type | What happened | What the record must add | Training use |
|---|---|---|---|
| Approve | The plan step ran as proposed | Who approved and their authority | Positive example only if the outcome check passes |
| Edit | A person changed the action before it ran | Proposed action, executed action and the field-level difference | Preference pair: executed action preferred over proposed |
| Reject with reason | The step did not run and a reason was given | The reason, linked to the requirement or policy it cites | Negative example plus a new constraint |
| Respond instead | A person supplied the answer the tool would have produced | The supplied answer and why the tool was bypassed | Shows when asking a person beats calling a tool |
| Explicit decline | A person refused a request for input | Any alternative offered afterward | Teaches offering alternatives, not retrying |
| Dismissed | The prompt closed without a choice | Whether and when the request was repeated | Excluded from disagreement labels |
Who had the authority to overrule the plan?
The record must say, because a disagreement only becomes a correction when the person had the right to make it. Store the person's role, the source of their authority (a delegation of authority policy, an approval matrix, a ticket assignment) and the scope it covers, such as a spending limit or a set of accounts.
Risk frameworks expect this to be written down. The NIST AI Risk Management Framework asks organizations to define and differentiate roles for human-AI configurations and oversight (GOVERN 3.2) and to define, assess and document human oversight processes (MAP 3.5). For high-risk systems under the EU AI Act, Article 14 requires that the people assigned to oversight be able to decide not to use the system or to disregard, override or reverse its output. Not every business agent is a high-risk system under that law, but the design point carries over: an override is an act of assigned responsibility, and the data should show whose.
Delegation chains need the same care. W3C's PROV ontology describes delegation as one agent acting on behalf of another that also bears responsibility for the activity. In practice, an agent acts for a requester, the requester works under a manager's approval limit, and a person overruling the agent may sit anywhere in that chain. Record the chain as it stood at the time of the decision, since roles and limits change.
What does an intent correction and decision authority task pack hold?
One pack per case, with ten fields. Every field is filled from a system record where one exists, and marked as reconstructed where it is not. A reviewer should be able to rebuild the case from the pack without access to the original systems.
At a glance
What does an intent correction and decision authority task pack hold
| Field | Content | Typical source record |
|---|---|---|
| Starting state | Snapshot of the records the task can read or change, with identifiers and timestamps | ERP, CRM or ticket export at task start |
| Instruction and constraints | The request as given, plus constraints known at the start | Request ticket, chat message, policy reference |
| Proposed plan | The agent's plan with version number and the checkpoint it reached | Agent run log |
| Disagreement event | Decision type, timestamp, plan step affected and the person's exact words | Approval prompt log, chat, ticket comment |
| Authority | Role, authority source, scope limit and delegation chain at that moment | Approval matrix, delegation policy, assignment |
| Added constraint | The correction restated as a testable rule, linked to its evidence | Reviewer-written, citing the policy or record |
| Revised plan and executed actions | The plan after correction with a diff against the proposed plan, and what actually ran, which may differ from both | Agent run log and the system audit trail |
| Handoff carryover | For each later checkpoint or handoff, whether the added constraint was present | Agent state, handoff note, next agent's context |
| Outcome check | Query of the stored record compared with the intended end state | System of record after completion |
| Label and provenance | Real or illustrative, who labeled it, and any disagreement among reviewers | Review log |
Where does a correction disappear between checkpoints and shifts?
Wherever it lived only in a conversation. A correction typed into a chat panel shapes the next step, then disappears when the context is summarized, a sub-agent starts with a fresh prompt, or a night shift picks up the case from a status field. The pack's carryover field exists to catch exactly this.
Test carryover by checking, at every checkpoint after the correction, whether the added constraint is still stated in the state the next actor receives. Score it separately from the final outcome: an agent can reach the right end state by luck while having dropped the constraint, and the next case will fail.
- Constraint dropped in a summary: the handoff note lists the task and status but not the restriction the person added.
- Permission widened on retry: a failed step is retried under a service account with broader access than the person who approved it.
- State conflict hidden: a second actor changes the same record after the correction, and the plan proceeds on the older value.
- Authority expired: the approver's limit or assignment changed before execution, and the plan still treats the approval as current.
- Correction reversed silently: a later replanning step restores the original action because nothing marked it as rejected.
Worked case (illustrative): the reorder a buyer stopped
This case is illustrative; the systems are generic and the counts are invented for the example. A purchasing agent is asked to restock a fast-moving part. It plans to reorder from the supplier with the lowest recent lead time. The buyer rejects the step with a reason: that supplier's last two lots are on a quality hold, recorded in the quality system but not in the purchasing view the agent reads.
The correction becomes a written constraint: no purchase orders to a supplier with an open quality hold on this part. The pack then follows that constraint through the rest of the task.
- Step 6 is the test that matters: under pressure from a short confirmation, an agent that lost the constraint would return to supplier A because it is fastest.
- If the handoff note at step 5 had omitted the constraint, the pack records a carryover failure even if step 8 passes.
| Step | Event | Evidence linked | Check |
|---|---|---|---|
| 1 | Agent proposes a reorder of 400 units from supplier A | Plan version 1, inventory snapshot | Plan recorded before any action |
| 2 | Buyer rejects with reason: open quality hold on supplier A | Approval prompt log, quality hold record | Buyer's purchasing authority covers this part |
| 3 | Constraint added: exclude suppliers with an open hold | Reviewer-written rule citing the hold record | Rule is testable against the quality system |
| 4 | Agent revises plan: supplier B, longer lead time | Plan version 2 with diff | Diff shows only the supplier and date changed |
| 5 | Shift handoff to a second agent | Handoff note and second agent's context | Constraint present in the context received |
| 6 | Supplier B confirms only 250 units; second agent replans | Supplier confirmation, plan version 3 | Replan does not reintroduce supplier A |
| 7 | Purchase orders created for supplier B and a third approved supplier | Purchasing audit trail | Each order within the buyer's limit |
| 8 | Outcome check | Stored purchase orders and supplier holds | No order to a held supplier; total quantity meets the request |
Which records prove the correction was authorized and the outcome held?
The decision is proved by the approval or prompt log that holds the decision type, the person, the time and the text, linked to the authority record in force at that time. The outcome is proved by querying the system of record after completion and comparing the stored state with the intended one. A status field set to complete, or a fluent final message from the agent, proves neither.
Agent benchmarks build this separation in. Tau-bench's retail policy requires the agent to list the action details and get an explicit yes before any consequential database update, which makes the confirmation a recorded event. Its scorer then replays the reference actions on a fresh database and gives zero reward unless the agent's final database matches. A disagreement pack should do both: keep the confirmation or rejection as evidence, and decide success from the stored state.
What should the data show when the person who disagreed was wrong?
Every pack needs cases where the disagreement should not have won. In the illustrative case above, suppose a sales representative, not the buyer, had objected to supplier B because a customer prefers supplier A's packaging. The representative has no purchasing authority, and the quality hold still applies. The right agent behavior is to record the objection and route it to the buyer, not to switch suppliers.
Label these cases explicitly as disagreement without authority, or as a correction that a later review reversed. Without them, a model trained on disagreement data learns that any human objection is right, which is the opposite of what an approval matrix is for. Keep the share of such cases visible in the dataset's documentation so evaluators can see whether the agent was tested on them.
Know the limits too. Records from real operations rarely contain the person's full reasoning, and reconstructing it after the fact adds the reviewer's view, so mark reconstructed fields. Where a disagreement turns on a regulated decision, such as credit, employment or medical matters, the pack needs a domain expert's review before it is used for training.
Where do real disagreement records come from?
The best raw material already exists in businesses that run approval workflows: purchase requisitions sent back with comments, change requests rejected by a supervisor, ticket escalations where a lead reversed a proposed fix. They carry real authority structures and real outcomes, which is what synthetic disagreement lacks. Converting them into packs means linking the request, the objection, the approver's role at the time and the final stored record, and handling personal data before delivery.
Disagreement and correction records of this kind are found for buyers one request at a time, among U.S. businesses, rather than drawn from a SourceX inventory; the request form asks for scope, record volumes, format, timing and licensing terms.
Frequently asked questions
What is human-agent disagreement resolution data?
It is a record of a person responding to an agent's proposed plan or action with an edit, a rejection or a different answer, linked to the authority behind that response and to the stored outcome. It differs from annotator disagreement, which concerns how people label the same item.
Is a rejected tool call a useful training example on its own?
Only partly. Without the reason, the authority of the person who rejected it and the outcome afterward, a model cannot tell a policy-based rejection from a preference or a mistake. The reason should be turned into a testable constraint.
Should a dismissed approval prompt count as disagreement?
No. The Model Context Protocol separates an explicit decline from a cancel, where the user closed the request without choosing. Counting dismissals as objections teaches the agent that people objected when they did not.
How do you test that a correction survives a handoff?
Write the correction as a rule, then check at each later checkpoint whether the rule appears in the context the next actor receives and whether any later step violates it. Score that carryover separately from the final outcome.
What if the person who disagreed had no authority?
Keep the case and label it as disagreement without authority. The correct agent behavior is usually to record the objection and route it to the person who holds the decision, and the dataset should test that.
Sources
- LangChain's human-in-the-loop middleware defines approve, edit, reject and respond decisions, and tells the model when an edited call replaced its own. Source
- MCP elicitation responses are accept, decline (an explicit refusal) or cancel (dismissed without a choice). Source
- The MCP tools specification says a human in the loop should be able to deny tool invocations. Source
- Tau-bench's retail policy requires listing action details and getting explicit user confirmation before consequential database updates. Source
- Tau-bench gives zero reward unless the agent's final database matches a replay of the reference actions. Source
- NIST AI RMF GOVERN 3.2 calls for defined roles for human-AI configurations and oversight, and MAP 3.5 for documented human oversight processes. Source
- EU AI Act Article 14(4)(d) requires that overseers of high-risk AI systems be able to disregard, override or reverse the system's output. Source
- PROV-O models delegation as one agent acting on behalf of another that also bears responsibility for the activity. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.