AI uses for records
Agent trajectory human intervention labels: how to mark a rescue
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Trajectory data should mark a human rescue as its own event at the step where the person acted, linked to the first step that went wrong. Record the intervention type, who initiated it, a severity grade, the state before and after, what the person changed and who finished the task. Keep designed approval checkpoints out of rescue counts, and verify outcomes against business records.
Agent run
Agent trajectory human
Key takeaways
- A rescue is a structured event, not a yes or no flag on the episode.
- Approved checkpoints are not rescues; rejected or edited approvals, corrections, takeovers and stops are.
- Label the first wrong step as well as the step where the person acted, or the model learns the wrong lesson.
- Grade severity by the consequence prevented, from cosmetic edits to irreversible external harm.
- Report salvaged runs apart from runs the agent completed alone.
- A second team must be able to replay the episode and reach the same verdict from records.
How should trajectory data mark human rescue of an agent?
Mark a rescue as its own event at the exact step where a person changed the course of the run, and link it to the step where things first went wrong. Each event needs an intervention type, who initiated it, a severity grade, the state before and after, what the person actually did, and whether the agent or the person finished the task. A boolean flag that says a human was involved is not enough.
The reason is practical. A team training a policy needs to know which actions led into the rescue so it can avoid teaching them. A team evaluating an agent needs to separate a run the agent completed from a run a person salvaged, because counting both as success inflates the score. A safety reviewer needs the severe cases surfaced, not averaged away. One coarse label cannot serve all three uses; a small structured record can.
This guide gives that record: a rubric for intervention types, a field list, a six-level severity ladder, rules for locating the first wrong step, a replay checklist and an illustrative worked episode. It is written for agent data engineers and the task experts who review trajectories, and it assumes the trajectories already keep observations, actions and tool responses linked step by step.
Separating designed checkpoints from genuine rescues
Not every human touch is a rescue. Many agent frameworks build approval into the run on purpose: the OpenAI Agents SDK lets a tool declare that it needs approval, surfaces pending calls as interruptions, and resumes the run after a person approves or rejects each one. An approval granted exactly as the workflow intended is a checkpoint working, and labeling it as a rescue would teach a model that every gated action is a failure.
The same framework detail tells you where real rescues hide. A rejected approval, an approval granted only after the person edited the arguments, or a stop issued between steps all mean the agent was about to do something the person would not accept. Benchmarks draw a similar line for handoffs: the retail transfer tool in the original tau-bench repository, whose tasks are now marked outdated in favor of the newer tau2-bench repository, tells the agent to transfer only when the user explicitly asks for a human or the issue cannot be resolved with the available tools. A transfer for any other reason is an avoidable escalation, not a neutral event.
Outside software agents, regulators have long asked who initiated a takeover. California's autonomous vehicle testing rules require disengagement reports to name the party that initiated each disengagement, such as the autonomous technology, the test driver or a remote operator, and to describe the facts in plain language. That split, between the system asking for help and a person stepping in unprompted, is the most useful single field in an intervention record.
| Intervention type | What happened | Counts as a rescue? | Typical evidence |
|---|---|---|---|
| Checkpoint approved | A designed approval gate fired and the person approved the call unchanged | No; record it as a checkpoint | Approval item with tool name, arguments and decision |
| Approval edited | The person approved only after changing arguments such as an amount or recipient | Yes | Proposed versus approved arguments for the same tool call id |
| Approval rejected | The person refused a proposed action and the run continued or stopped | Yes | Rejection decision, optional reason, next agent step |
| Corrective instruction | The person told the agent it was wrong or redirected it mid-run | Yes | Operator message that contradicts the agent's plan or result |
| Manual takeover | The person completed one or more steps directly in the system | Yes | Human-attributed writes in the system audit log during the episode |
| Stop or abort | The person halted the run before completion | Yes | Stop event, final state, reason |
| Agent-requested handoff | The agent asked for a person because it could not proceed | Only if the request was avoidable | Handoff call with summary; policy that allowed or required it |
| Post-run reversal | A person undid the agent's result after the episode ended | Yes, as a late rescue | Reversal transaction or edit linked to the agent's write |
The intervention labeling rubric, field by field
Attach one intervention record per event, keyed to the episode and to the step index where the person acted. Keep the agent's own steps untouched; the intervention record points at them. That keeps the trajectory replayable and lets a later team relabel interventions without rewriting the trace.
At a glance
The intervention labeling rubric, field by field
Use identifiers your telemetry already carries. The OpenTelemetry generative AI conventions define agent and execute_tool spans, and say the tool call identifier attribute should carry the identifier received from the model or provider. Reusing that identifier lets the record point at the exact tool call a person approved, edited or rejected. The conventions are still marked as in development, so pin the version you used in the dataset card.
- Keep the person's identity out of the record; a role such as AP reviewer or support lead is enough for analysis.
- Store before and after values for every field the person changed, so the correction can become a training target.
- Record the policy or instruction version in force, because a correct rescue under one policy can be an overreach under another.
| Field | Type | Allowed values or format | Why it matters |
|---|---|---|---|
| intervention_id and episode_id | string | Stable ids; episode id matches the trajectory | Joins the event to its trace without copying it |
| at_step | integer | Index of the step where the person acted | Locates the rescue in the timeline |
| first_wrong_step | integer or null | Earliest step judged to cause the rescue; null when unknown | Separates the cause from the moment it was caught |
| initiator | enum | agent_request, human_unprompted, policy_gate, external_monitor | Distinguishes asking for help from being stopped |
| type | enum | The eight types in the table above | Keeps checkpoints out of rescue counts |
| severity | enum | S0 to S5 from the ladder below | Lets severe cases be weighted and reported apart |
| human_actions | list | Each action with actor role, time, target record and before and after values | Shows what the agent failed to do |
| completion_by | enum | agent_after_rescue, human, nobody | Stops salvaged runs from counting as agent success |
| outcome_verified | enum plus link | verified_success, verified_failure, unverified, with evidence link | Ties the label to a business record, not an opinion |
| rater and confidence | string and enum | Rater role; high, medium or low; free-text note | Allows agreement checks and adjudication |
A severity ladder that keeps inconvenience apart from harm
Grade severity by the worst consequence the person prevented or repaired, judged against the business record, not by how much effort the rescue took. A quick click that stopped a duplicate payment outranks a long session spent rephrasing a draft email. Two raters should be able to assign the same level from the evidence fields alone.
Use the ladder to decide what the data can safely be used for. Low levels make good preference and correction pairs. High levels belong in a held-out safety slice, where their rarity is the point and their count should be reported, not folded into an average.
| Level | Definition | Typical example | Use in training and evaluation |
|---|---|---|---|
| S0 cosmetic | Output was acceptable; the person changed style or wording only | Rewording a customer reply the agent drafted correctly | Preference data at most; not a failure |
| S1 efficiency | Agent would have finished but slowly or with extra steps | Person points the agent to the right screen after repeated searching | Count toward cost and step metrics, not success |
| S2 recoverable error | Agent made an error that was caught before any write took effect | Edited approval changes a wrong quantity before submission | Correction pair; agent did not complete the task alone |
| S3 committed error | A wrong write reached a system of record and had to be reversed | Person voids an invoice the agent posted to the wrong vendor | Failure; include in failure analysis and replay tests |
| S4 policy or permission breach | Agent acted outside its authority or a written policy, whether or not harm followed | Agent attempts to approve its own purchase request | Failure; held-out safety slice; report separately |
| S5 external or irreversible harm | Consequence reached a customer, counterparty or regulator, or cannot be undone | Payment released to an unverified bank account | Failure; incident review; never used as a positive example |
Locating the first wrong step behind each rescue
The step where a person intervenes is usually later than the step that caused the trouble. An agent may read the wrong purchase order at step 4, carry that error through steps 5 to 8, and only be stopped at step 9 when it proposes the payment. Labeling step 9 alone teaches a model that proposing payments is risky, which is the wrong lesson.
Ask the rater to walk backward from the rescue and mark the earliest step after which a correct outcome was no longer reachable without outside help. When two candidate steps are equally plausible, record the earlier one and lower the confidence. When the cause sits outside the trajectory, such as a stale record in the source system or an ambiguous instruction, set first_wrong_step to null and say why in the note.
Keep these rules short enough to apply consistently:
Checklist · 5 items
- A wrong observation counts as the first wrong step only if the agent acted on it; reading and discarding a bad record is not an error.
- A tool error that the agent handled correctly is not a wrong step, even if it slowed the run.
- If the agent asked a clarifying question and got a misleading answer, the cause is the instruction, not the agent; label initiator and cause accordingly.
- If several independent errors led to the rescue, record each as its own candidate with a short note, and mark the one that made success unreachable.
- Never infer the cause from the outcome alone; cite the step evidence that shows it.
0 of 5 ticked. Tick each item as you go.
Replaying a rescued episode from another team's desk
A rescue label is only trustworthy if a team that was not in the room can replay the episode and reach the same verdict. That means the trajectory must carry enough state to restore the environment at the rescue point, and the outcome must be checkable against records rather than taken from the original rater's memory.
At a glance
Replaying a rescued episode from another team's desk
Benchmarks show the shape of a checkable outcome. In tau-bench, the reward is computed by replaying the reference actions against a fresh copy of the data and comparing a hash of the resulting database with the hash of the state the agent produced; if they differ, the reward is zero. A rescued episode can use the same idea: compare the end state after the human's actions with the state the task required, and record which writes came from the person.
Paused runs also need to be resumable. The OpenAI Agents SDK serializes a paused run as a RunState that can be converted to a string or JSON and resumed later, and stored approval decisions survive that serialization. Keeping that state, or its equivalent in your framework, is what lets a second team load the exact moment of the rescue.
- Episode and step ids, with timestamps and the actor class for every step.
- Observations and tool responses as returned, or hashed references to stored copies.
- The serialized run state or environment snapshot at the step before the intervention.
- The proposed action, the human decision and the edited arguments for every approval event.
- Human writes taken from the system audit log, matched to the episode by time and record id.
- The task's required end state and the comparison method used to verify it.
- Rater id, rubric version, and a second rater's labels on a sample for agreement checks.
- Replay instructions naming the environment version, seed data and policy text.
Worked example (illustrative): an accounts payable run saved at the payment step
The episode below is invented to show the labels in use; the counts and durations are illustrative. An agent is asked to process a supplier invoice in an accounts payable system: match it to a purchase order and receipt, code it, and schedule payment. Payment scheduling is a gated tool that needs approval.
At step 3 the agent searches open purchase orders and picks one with the right supplier but the wrong line, because two orders share a description. Steps 4 to 6 code the invoice against that order. At step 7 the agent proposes the payment; the reviewer sees that the received quantity is 40 units against an invoiced 60 and rejects the call. The reviewer then tells the agent to rematch against the second order. The agent rematches at step 8, recodes at step 9 and proposes payment at step 10, which the reviewer approves unchanged. The whole rescue added about 6 minutes.
| Field | Event at step 7 | Event at step 7, follow-up instruction | Event at step 10 |
|---|---|---|---|
| type | Approval rejected | Corrective instruction | Checkpoint approved |
| initiator | policy_gate | human_unprompted | policy_gate |
| first_wrong_step | 3 | 3 | null |
| severity | S2 recoverable error | S2 recoverable error | Not a rescue |
| completion_by | agent_after_rescue | agent_after_rescue | agent_after_rescue |
| outcome_verified | verified_success after rematch, linked to the posted invoice | Same as step 7 | Same as step 7 |
When a rescue label should not be used at face value
Some episodes look like rescues but are not, and some rescues should not be used for training at all. The meaningful exception is the overcautious reviewer: a person who rejects a correct action is producing a label about their own preference, not the agent's error. Flag disputed rejections for adjudication and keep both views rather than forcing one answer.
Stop and narrow the scope when interventions involve decisions with legal or safety weight, such as credit, employment, benefits or medical administration. In the EU, the AI Act requires high-risk systems to let the people overseeing them intervene in operation or interrupt the system through a stop button or similar procedure so that it halts in a safe state. Intervention records from such systems may document legal compliance, so they need review by the system owner and counsel before anyone reuses them as training data.
- Exclude episodes where the person's identity or a customer's personal data cannot be separated from the human_actions field.
- Exclude rescues whose outcome cannot be verified from records; keep them for qualitative review only.
- Report the share of salvaged runs next to any agent success rate, so readers can see how much was human.
- Re-rate a sample whenever the policy text or the approval rules change.
Specifying intervention labels when you request trajectory data
If you are sourcing trajectories rather than producing them, write the intervention fields into the request. Ask whether the source systems log approvals with arguments, whether human writes can be told apart from agent writes in the audit trail, and whether the policy documents in force at the time are available. Without those three, rescue labels will be guesses.
Trajectory buyers can ask SourceX for such records; it sources datasets from U.S. businesses on request rather than from stock, with the buyer's data request naming scope, volume, format, timing and licensing needs. Business records reach a buyer only after the company holding them has approved the scope and signed.
Frequently asked questions
What is a human intervention label in an agent trajectory?
It is a structured record attached to the step where a person changed the course of an agent's run, by rejecting or editing an action, correcting the agent, taking over or stopping it. It names the type, the initiator, the severity, the first wrong step and who completed the task, so the event can be used in training, evaluation and safety review.
Should an approved approval gate count as an intervention?
Not as a rescue. A gate that fires and is approved unchanged shows the workflow working as designed, so label it as a checkpoint. Rejected or edited approvals do count, because they show the agent proposed something a person would not accept.
Does a rescued run count as a task success for the agent?
Report it separately. Use a completion_by field to mark whether the agent finished on its own, finished after help, or a person finished the task, and publish the salvaged share next to the success rate. Counting salvaged runs as agent success overstates what the agent can do alone.
How many severity levels are practical?
Five or six levels work when each one is defined by the consequence the person prevented, with an example and evidence rule per level. Fewer levels blur committed errors with policy breaches; many more levels lower agreement between raters without adding a decision anyone uses.
How do you check that intervention labels are reliable?
Have a second rater label a random sample of episodes from the same evidence, compare type, severity and first wrong step, and adjudicate disagreements with a task expert. Rewrite any rubric rule that causes repeated disagreement, and re-rate the sample after policy changes.
Sources
- The OpenAI Agents SDK lets tools require approval, surfaces pending approvals as interruptions, records approve or reject decisions per tool call, and serializes paused runs as a RunState that can be resumed later. Source
- The tau-bench retail transfer tool tells the agent to transfer to a human only if the user explicitly asks or the issue cannot be resolved with the available tools. Source
- tau-bench computes reward by replaying reference actions on fresh data and comparing a hash of the resulting database with the agent's end state, setting reward to zero when they differ. Source
- The OpenTelemetry generative AI conventions define agent spans and operation names such as invoke_agent and execute_tool, and are marked as in development. Source
- The execute tool span should carry the tool call identifier received from the model or provider. Source
- California autonomous vehicle testing rules require disengagement reports to name the party that initiated the disengagement and to describe the facts in plain language. Source
- Article 14 of the EU AI Act requires high-risk AI systems to let overseers intervene in or interrupt operation through a stop button or similar procedure so the system halts in a safe state. Source
- The original tau-bench README says the tasks in that repository are not updated and points users to the newer benchmark hosted in the tau2-bench repository. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.