Use Neatlogs without code
Read agent runs, report problems, review quality, and follow resolutions without owning instrumentation.
Use Neatlogs without code
Read runs, flag problems, and work with engineers. No code required.
You do not need to understand instrumentation or inspect raw span payloads. Neatlogs connects the agent's output to the evidence, quality decisions, investigation, and resolution.
What you can do
| Task | What you contribute |
|---|---|
| Read a run | Confirm what the customer asked and what the agent did |
| Flag a problem | Describe the mismatch and expected behavior |
| Add context | Supply policy, customer, or business facts telemetry cannot infer |
| Review an Eval | Apply a defined quality standard |
| Help an investigation | Answer focused questions and challenge unsupported conclusions |
| Follow an incident | Track what changed and whether the problem returned |
Engineering can continue to own instrumentation, code changes, tests, and deployment.
Follow one failure without reading raw telemetry
Suppose a customer reports that the refund assistant offered an expired policy.
1. Find the interaction
In Traces, search by customer, session, time, workflow, or a phrase from the report. Open the candidate run and confirm that the first input and final output match the customer interaction.
You do not need to inspect every span. Start with the rendered input and output, then follow the relevant model, tool, or retrieval steps only when needed.
2. Flag the exact problem
Highlight the sentence or action that was wrong and add a comment to it. You can also open Comments on the trace, or use the Comment action on one step.
Observed: The answer offered a refund under policy v2.
Expected: Use the policy active on the order date.
Impact: The customer received an option support cannot honor.
Unknown: Why policy v2 was selected.Keep the cause under Unknown unless the evidence already proves it.
3. Add business context
Attach the policy, expected outcome, customer impact, and any exception another reviewer would need.
Telemetry can show which policy the agent retrieved. It cannot decide which policy was valid unless that rule exists in the trace or you provide it.
4. Start or support the investigation
Start an investigation from the incident with Investigate, or select Ask neat AI on the trace and ask it to investigate the run. The investigation reads the linked evidence and may compare failed and successful runs. If neat AI asks a focused question about a business fact, answer it or provide the authoritative source.
Open important citations in the final report. Challenge a conclusion when:
- it assumes a policy that was not supplied;
- it confuses the visible symptom with the cause;
- the cited evidence does not support the claim;
- an alternative explanation still fits.
5. Follow the resolution
The reviewed incident can be handed to an engineer, a tracker such as Linear or Jira, or a configured coding agent. You can follow the evidence, incident, ticket, and verification without approving code or deployment.
After release, check whether the original behavior stopped and whether the customer-facing outcome is now correct.
Review quality through Evals
When an Eval is assigned to you, open Evals and select Assigned to me. Then:
- read the question and verdict definitions;
- confirm whether you are judging the whole run or one step;
- inspect only the evidence needed;
- choose the configured verdict;
- add reasoning for ambiguous or policy-dependent cases.
If the Eval offers a verdict such as Cannot determine or Insufficient evidence, use it when appropriate. Do not guess because the form expects an answer.
The web review queue is the canonical surface. Slack-based review depends on the interactive workflows enabled for the workspace.
Ask questions in plain language
Ask in Investigate for questions across the project, or select Ask neat AI on a trace for questions about that run. Useful questions focus on evidence:
- Which policy document was used in the final answer?
- What did the order lookup tool return?
- Where did the account status first appear?
- Did the agent retry the same tool?
- Show successful runs from the same workflow for comparison.
Open the linked traces behind important answers. Plain-language search helps navigate evidence, but it does not turn an uncited explanation into fact.
What remains permission-controlled
Using Neatlogs without code does not grant authority to:
- change instrumentation;
- access every project;
- expose connected tools to an evaluator;
- dispatch work to every repository;
- approve a pull request;
- deploy or roll back a change.
Those actions depend on role, project, external integration, repository, and deployment permissions. Reading runs and commenting work with any project role, including Project Viewer.
Frequently asked questions
Do I need to read raw traces?
No. Use rendered inputs, outputs, execution views, and plain-language questions. Raw detail remains available when a specialist needs it.
Do I need to write code?
No. Engineering can own instrumentation and implementation.
What can I do from Slack?
Slack can receive configured alerts. Interactive Eval or investigation responses depend on workspace-specific capabilities.
Can I approve or deploy a fix?
Only if your role and the external workflow grant that authority. Reviewing evidence does not grant code or deployment access.
Do I need a Neatlogs account?
Dashboard collaboration normally requires project access. An enabled guest review flow may allow a narrower task without full membership.