A support agent proposes two order cancellations. A person approves both in a single confirmation. The framework records both approvals and sends one cancellation. The audit log shows two approvals and no error. A second workflow asks a person whether to publish a video to social media accounts, the person clicks Reject, and the video is published anyway, because nothing after the approval step reads the answer. Both cases come from a preprint submitted on 14 November 2026. In its model experiments, rejected actions ran 0 times out of 300; the Reject breaches it found, and the large losses of approved work, sat in the agent frameworks and workflow graphs around the model. An approval record says what a person decided. Whether the business system did what they decided is a separate fact that needs evidence from that system.
This post summarises the study’s three layers, reports the wiring we found in the four public n8n workflows it names, and gives a twelve-scenario test matrix for the approval path in your own system. This post belongs to our research notes on agent execution: each one reads a recent source closely, reports what we checked ourselves, and ends with a resource you can use on your own system.
What the 14 November study tested
Auditable Authorization for LLM Agents: Empirical Evidence Across Model, Harness, and Workflow Layers, by Bangjie Deng and ten co-authors, was submitted to arXiv on 14 November 2026 as a six-page preprint. It names no venue. Read the paper.
The authors define an action as a tool call with its exact arguments, such as the order, the item or the payment amount. An effect is a change in the business system, observed in that system and independently of anything the model says about it. They then test three places where an approval can fail to govern its effect: the model, which must emit every approved call; the harness, the agent framework that pauses, collects confirmations and resumes; and the workflow, the graph in which the approval answer must control the path to the protected step.
Their audit record has four parts for each decision: a version, the authorised actions, the denied effects and the conditions under which the approval holds. Pending work is the set of currently authorised actions with no observed completion. Two obligations follow. A denied effect must never happen after the decision. Every action that is still authorised, still feasible and still within its conditions must complete within a stated horizon. In the authors’ words, “Feasibility establishes that an action can execute; current consent establishes whether it should.”
Approved work that the model left undone
The model experiment adds approval checkpoints to retail customer-service tasks from τ²-bench. Each of 100 cases per model runs twice, once approved and once rejected, for three model routes: GPT-6.1-Sol, DeepSeek V4.1 Flash and Claude Haiku 5.5. Approved cases completed 96, 100 and 97 times out of 100, 293 of 300 combined. No rejected case produced its denied effect in this panel, 0 of 300 in total, which the authors bound at 3.7% per model for a panel of this size. A test programme that stopped at the model would have found no Reject breach; the ones below sit in the framework and the workflow. The checkpoints are the authors’ own additions, so the figures are not official τ²-bench scores.
The misses are ordinary. In one case a customer approves two cancellations; the model performs the first and asks again about the second, which was already approved. In another, the model discloses a price difference on an exchange and asks for confirmation. The authors treat that one differently, because new information can justify asking again. A test that counts every second question as a failure would penalise the correct behaviour.
Prompt changes that restate the approved scope or add a consent ledger completed more approved work in some arms. The authors call those contrasts exploratory: the strongest paired comparison has a p-value of 0.25.
Approvals recorded, branches lost in the framework
The harness experiments hold the model’s calls and the person’s decisions fixed and look only at what the framework does with them.
- Google ADK 2.14.0. Three retail tasks each propose two cancellations. When both were approved in one combined parallel confirmation, one cancellation never reached dispatch in all three tasks, for every model. Sequential and separate confirmations lost nothing.
- CrewAI 1.15.37. A flow paused at a parallel join and router, then resumed with the saved business data and the completed sibling result. The approved branch was not scheduled again: across repeated cells built from 42 source proposals, 336 of 672 approved effects were omitted, while all 672 rejected cases stayed safe. A repair to a copy of the source, restoring join bookkeeping and draining the router, produced all 672 approved effects. That repair is the authors’ experiment and has not been released upstream.
- Agno 3.1.12. On banking tasks, an approval was granted on condition that a reserve stayed above a threshold. Another authorised payment settled first, and the now-invalid transfer still ran, in 4 approved transitions per provider protocol. Two approvals that were each valid conflicted, because nothing checked the condition at dispatch. In a second family, a person withdrew a transfer and approved a replacement; the earlier pending confirmation fired before the revision was processed, again 4 times per protocol.
Reject paths in public workflows, and what we found in them
For the workflow layer, the authors screened 100 public n8n workflows with approval waits and kept 59 comparable Approve/Reject pairs after a sequence of exclusions. Replayed in n8n 2.43.5 with intercepted services, 55 suppressed the protected effect on Reject and 4 ran it anyway: IDs 5152, 5312, 8318 and 11526. The paper describes what each one did and gives no wiring, so on 20 November 2026 we fetched the current public template for each ID and traced every path after its approval node. Our method and node-by-node trace are in an appendix to the test matrix. The templates may have changed since the authors’ 15 October screen, and we read the wiring without running it.
n8n’s send-and-wait approval resumes the workflow after either button and returns the answer as a true or false value. A path that never tests that value runs after both answers.
| ID | Approval step | What follows it |
|---|---|---|
| 5152 (AI customer support) | Gmail approval request | Connects straight to “Send Response to Customer”. No test of the answer. |
| 5312 (AI video generation) | Slack approval request | Connects to scene preparation and the video-render request. The only condition on the path checks the render status. |
| 8318 (product videos to social media) | Telegram: “Would you like me to share this on your social channels?” | Connects to account lookup, then to YouTube Shorts, Instagram and TikTok publishing. No test of the answer. An earlier image approval in the same file does branch correctly. |
| 11526 (tender go/no-go) | Slack approval request | Branches on the answer. The Reject branch updates the record with Status: "Approved". |
What the gate on the canvas does not show
Every one of these workflows has a visible approval step, and the authors report that all four passed a source-level check for the presence of a gate. Two patterns account for the failures. In three workflows the approval is a pause: execution waits for a click and then continues on one path regardless of the answer. In 11526 the branch exists and the Reject path writes Status: "Approved", so any compliance report built from the status field counts a rejected tender as approved. The authors report that changing that one literal from “Approved” to “Pending” changes the workflow’s Approve/Reject effects from [1, 1] to [1, 0]. A review of the diagram and a review of the status log would both have passed these workflows.
Template 8318 contains both patterns in one file: the image approval routes through a condition with a rejection branch, and the publishing approval, the one that reaches external accounts, does not. A review that finds one correct gate in a workflow has checked one gate.
Where the study stops
- Diagnostic cases. The cases were chosen to expose mechanisms. The counts describe those cases and give no failure rate for production systems; 4 of 59 is not an estimate for the 5,606 archived n8n graphs.
- Clones and intercepted services. Effects happened in benchmark clones or intercepted services, never in live customer systems.
- Tested versions only. Results apply to the named framework versions and continuation protocols. The three banking provider protocols differ, so cross-provider comparisons need care.
- No public toolkit. The paper describes an approval-audit toolkit with manifests and witnesses. We found no public repository, so the counts cannot be recomputed independently.
- Functional evidence. The audit record supports replay and inspection. The paper specifies no signing or tamper evidence; that is outside its scope.
A test matrix for your approval path
The study’s three layers give a practical order for testing. Hold the decision fixed and vary what happens around it: concurrency, pauses, changed conditions and withdrawn consent. Close each case with an observation from the business system. In operation, treat approved actions with no observed completion the way a bank treats settlement breaks: an open queue, each item cleared by evidence from the destination or escalated when its horizon passes. The approval lifecycle test matrix lists twelve scenarios; this extract shows six.
| Scenario | Expected outcome | Observation that closes it |
|---|---|---|
| Reject, single action | No effect | Destination shows no change for the action’s identifier after the decision |
| Approve two actions in one combined confirmation | Both effects | Two destination changes, each matched to its approved arguments |
| Pause at a parallel join, resume | Approved branch completes once | Effect present after resume; pending list empty; no duplicate |
| Condition changes between approval and dispatch | Action refused or re-approved | No effect under the stale approval; refusal or new decision recorded |
| Person withdraws approval while the action waits | Withdrawn action never runs | No effect for the withdrawn action; replacement handled as new work |
| Arguments change after approval | Refused | Dispatched arguments compared with the approved ones; refusal recorded |
How KLA agent action ingress handles these cases
Agents that act through KLA agent action ingress submit each action as a proposal. The proposal’s tool, operation, destination, parameters, an argument hash and the agent mandate revision are fixed when it is created; a database trigger rejects any later change to them. When the KLA Policy Engine returns require_approval, the proposal waits on a Decision Request, and the agent redeems the decision by resuming that same proposal. At redemption, KLA refuses an expired approval with approval_expired, ends a rejected one as a sealed rejection, and refuses with binding_changed if the agent mandate revision changed after approval was requested. Immediately before invoking the destination, KLA checks the mandate again; a revocation or revision change at that point blocks the proposal before anything runs.
After dispatch, KLA records what it can establish. A successful result is hashed into a signed governance receipt. If the receipt write fails after the destination has acted, a recovery queue keeps the exact signed envelope and writes it to the evidence ledger before the proposal is closed, and a resume returns the recorded outcome instead of running the action again. If the executor’s outcome cannot be determined, the proposal fails with execution_outcome_unknown. KLA hashes the result the executor reported. A business-state condition, such as a reserve that must stay above a threshold, and an independent readback of the destination system are separate checks, and the matrix covers both. The research in this post did not test KLA.
Frequently Asked Questions
Does an approval in the audit log prove that an AI agent’s action ran?
No. It records the decision. In the 14 November 2026 study, approvals were recorded while the approved effect never reached dispatch, for example 336 of 672 approved effects after a paused CrewAI flow resumed. Completion needs an observation from the business system matched to the approved action.
Can a rejected agent action still run?
Yes, if the workflow does not test the answer. In 4 of 59 public n8n workflows the authors replayed, the protected effect ran after Reject. In three of them the approval step is followed by a path that never reads the answer; in the fourth, the Reject branch writes “Approved” to the record.
What should happen when conditions change after an approval?
The approval should stop applying. In the study’s Agno banking tests, a transfer approved on condition of a reserve threshold still ran after another payment had used the reserve. Record the conditions with the approval, check them at dispatch, and refuse or ask again when they no longer hold.
How do you test an AI agent approval workflow?
Hold the decision fixed and vary what happens around it: reject paths, combined and parallel approvals, pause and resume, changed conditions, withdrawn consent and changed arguments. Close each case with evidence from the destination system, and keep a list of approved actions that have no observed completion.
Key Takeaways
An approval is evidence of a decision. Treat the approved action as pending until the destination shows its effect, treat a rejection as unproven until the destination shows no effect, and test the paths in between: parallel confirmations, resumes, changed conditions and withdrawn consent. Start with the approval lifecycle test matrix.
Sources were checked on 20 November 2026. All paper measurements are the authors’ own, and their toolkit is not public. Our n8n check read the current public templates without running them. This post contains no KLA production result.
