Your follow-up needs a follow-up.
Enquiries arrive. Ownership gets blurry. We map the handoffs, identify where work stalls and define a clearer next step.
Still copying updates, chasing handoffs and checking whether the automation actually worked? We help you find the friction, connect the workflow and check what happens next.
Business workflows · Integrations · Automated action checks
No AI yet? You’re in the right place.
Working solo or running a growing team? We start with how your business works today.
The little workarounds add up. We start with the part of your operation that keeps sending work back to people.
Enquiries arrive. Ownership gets blurry. We map the handoffs, identify where work stalls and define a clearer next step.
We review how your existing systems exchange information and scope practical integrations and automation around the work you need done.
For teams using agents and APIs, we scope checks for duplicate effects, changing permissions and the difference between a success log and the final state.
We look at your tools, the people using them and the work moving between them. Together, we choose a useful scope and agree what a better outcome would look like.
Let's look at your workflow ↗Map the current workflow, the bottleneck and the evidence available.
Agree the intervention, responsibilities, acceptance checks and scope.
Build or configure the agreed change, then check it in a suitable test environment.
Explain the findings, document the change and identify the next practical action.
An agent can choose the right tool, receive a success response and still leave a duplicate payment, a missing order or an unauthorised update. We investigate the gap between an authorised action and its persisted business effect.
Agent Transaction Integrity Test: one critical write, tested across failure boundaries, with the result checked in the destination system.
Before running a test, we agree what must remain true: the identity of the operation, the authorised destination, value or content, and the expected number of effects. “The workflow completed” is not the acceptance criterion. For an order, for example: one authorised request must correspond to one correct order record.
We target the boundaries between agent, orchestrator, tool and destination: lose the response after the destination accepts a write; retry or resume; race concurrent requests; change permission while work is waiting. Each scenario has an initial state, a controlled fault and an expected outcome.
We reconcile operation IDs and the relevant destination records, counts and values with the agreed invariant and authorisation. A tool response or agent trace is evidence of execution, not sufficient confirmation of the persisted effect. Where reads are delayed or evidence is missing, we retain an unconfirmed outcome instead of declaring success.
For each finding, we connect the scenario, permission context, attempts and external result. Your team receives the steps to reproduce, expected versus observed behaviour, supporting evidence and prioritised remediation guidance. We review existing controls so the recommendation addresses the actual failure boundary.
A narrow scope puts the work into one consequential action instead of a broad platform replacement. Controlled faults expose the recovery paths an ordinary successful run does not exercise. Business-specific checks turn “looks correct” into a checkable result. Correlated evidence gives your engineers a concrete sequence to reproduce and a control to investigate.
A payment is accepted, the response is lost, and the agent retries. The execution can finish while two charges exist. We correlate the logical operation with the destination records: expected effect count 1; observed count 2. The finding identifies the failing scenario and the controls to examine, rather than treating the completion status as proof.
Keep your observability, evaluation and control tools. The sprint brings them together around a transaction-level question: did this authorised action produce the intended effect under the conditions tested?
| Existing check | What it helps establish | What CPL verifies in the scoped test |
|---|---|---|
| Agent / tool evaluation | Tool choice, arguments, trajectory and configured quality or business checks. | The relationship between the authorised action, its attempts and the persisted effect in the destination. |
| Execution trace / success status | What the system recorded about calls and responses. | Whether the expected record, value and count are present after a lost response, retry or recovery. |
| Access and approval controls | Which actions are permitted and how decisions are recorded. | What happens when that permission changes while an action is waiting or being resumed. |
| Data reconciliation | Whether selected source and destination data match. | How controlled agent execution failures produce a mismatch, with a reproducible sequence and remediation priorities. |
Evaluation platforms can support custom business checks, and data-testing tools can reconcile systems. Our offer is the specialised investigation of your agreed flow, not a claim that those capabilities exist only at CPL.
Scenario and fault timing · expected invariant · attempts and operation IDs · relevant authorisation context · destination evidence · expected versus observed result · reproduction steps · existing-control review · prioritised remediation.
For engineering teams, automation agencies and platform owners who need an outside technical check on a specific integration.
Testing takes place in an agreed representative environment, with defined data and access boundaries. Production changes and ongoing monitoring are separate scopes.
If an error in your automation feels like a mystery, let’s talk. We investigate with enough precision that even Scooby-Doo and the gang might send us their CVs.
A synthetic study of automated action integrity.
Our controlled laboratory work used official Temporal and LangGraph components, with Promptfoo evaluating recorded evidence. The findings concern the tested applications and their controls.
A defined action, a controlled failure and an agreed rule for the expected outcome.
Compare recorded execution with observable downstream state and the controls in place.
Separate demonstrated findings, remaining uncertainty and practical remediation.
Published 18 October 2026 · Version 1.1 · DOI: 10.5281/zenodo.23436816 · Report CC BY 4.0; selected harness MIT.
Laboratory findings do not establish vendor vulnerabilities, a market-wide failure rate or production safety. Delivery scope is agreed for each engagement.
Workplace sketches, AI updates and the conversations your smart devices definitely have in our fictional universe. Created and performed with the same curiosity that drives the Lab.
Jade · Operations & deadlines
“Quick question. How long has this been a quick fix?”
The handoffs, the workarounds and the people keeping everything moving.
Judite · Legal & permissions
“Resolved by whom, with whose permission?”
Big announcements meet the person who reads the small print.
The AI · After-hours assistant
“I know the answer. I'm letting you appreciate the complexity.”
A fictional backstage life for the assistants who are always “working”.
Character concepts and sample lines. First episodes are in development.
Yes. You don’t need AI, an existing automation or a technical team to start the conversation. Bring the work that keeps eating your day — enquiries, follow-ups, bookings or admin. We begin with how you do it now and agree what would be useful to improve.
Owner. Admin. Sales. The person chasing the person who hasn’t replied. All you? You’re in the right place. We shape the scope around your workload and priorities, so you can see what’s worth changing and what it would cost before committing.
We begin with what you already use. Any recommendation to connect, configure or replace a tool should follow from the workflow and the agreed business objective.
Yes. Bring a handoff, repetitive task or automated action that needs attention. The initial discussion helps define whether an operational review, an integration or a technical verification sprint fits.
The integrity sprint is scoped for a representative test environment using agreed data and access boundaries. Production access and changes are outside the default test scope.
A short description of the workflow, the tools involved and what keeps going wrong. Keep customer records, passwords and production logs out of your first message.
Monitoring helps explain executions. This sprint tests an agreed business action under defined failure conditions and checks the resulting state in the destination. It can complement your existing monitoring and evaluation tools.
Yes. We can scope a focused outside check of one integration and critical action. We agree the environment, responsibilities and evidence before testing; implementation or remediation work is described separately.
Tell us what happens, where it gets stuck and what you want to change. We'll use that to frame a useful first conversation.
Tell Jade about it ↗Please keep your first message free of customer data and credentials.