One workflow. A clearer picture.
Find the failures.
Keep the tests.
AI Agent Reliability Sprint: a focused assessment for startups already building with LLMs. Turn important failures in one agent workflow into repeatable regression tests.
A direct conversation with me, the builder doing the work.
Offline demo / mocked LLM
$ python -m pytest
3 failedbuggy baseline
27 passedfixed test suite
Recorded 18 Oct 2026. Different suite sizes.
Sample on a demo system. Inspect the report ↗
01 / The scope
Know what you are buying.
Included
- Assess one agreed agent workflow.
- Reproduce its important failures using agreed examples.
- Add automated regression tests you can rerun.
- Validate tool-call structure and permitted tool selection.
- Deliver a documented test report, with findings and remaining risks.
Not included
- Indefinite maintenance or on-call support.
- A full agent or application rebuild.
- Additional workflows or open-ended feature development.
- A guarantee that an LLM will never fail.
02 / How it works
Agree the edges. Then get to work.
Plan for five working days after scope, access and examples are ready. I will confirm availability and dates in writing.
Define the workflow
Share the workflow, failure examples and what a correct result looks like. I confirm fit, deliverables, acceptance criteria and the fixed price before starting.
Reproduce and test
I capture the agreed failures, add regression coverage and implement the agreed tool-call validation. Findings outside the scope are recorded for a separate decision.
Review and hand over
You receive the tests, run instructions and report. We review the agreed acceptance criteria and any unresolved risks. Extra work needs a separate written scope.
03 / See the approach
Small demo.
Visible boundaries.
An offline, mocked agent illustrates the method. It is a demonstration project, with no client data or production claims.
See the demo method and run instructions ↗Sample on a demo system
Three failures worth catching.
- 01 / Malformed tool-call JSON Validate the envelope
- 02 / Wrong tool selected Check permission before dispatch
- 03 / Unsupported answer Require retrieved evidence
The repository keeps both the buggy and fixed implementations. The sample report states the execution status and limits of the evidence.
Read the sample report ↗Download the sample report (PDF) ↗
Fictional sample: example client handover report (HTML) ?
Fictional sample: example client handover report (PDF) ?
04 / About me
Henry Williams.
I am Henry Williams, a sole trader in London, trading as Sparrow Works. I have completed one real paid job using Webflow, Zapier and Airtable. My own projects include EvalForge.
I am early in my working life. These offers are deliberately bounded: a clear problem, agreed checks and a handover you can inspect. You work directly with me.
Start with the workflow
What is your agent getting wrong?
Send a short description, an example failure and your preferred timing. I will confirm whether the sprint fits.