How we work
Every Poush Digital engagement uses AI agents. They read, map and draft. The senior judgement stays human.
The result is more depth for the same budget. An assessment can cover the whole landscape instead of a sample, and the deliverables can stay current after the engagement ends.
What an agent harness is
An AI model on its own is a capable writer with no access to your world and no accountability. A harness is everything we put around it so it can do reliable work:
- Models. The AI models doing the reading and drafting, chosen for the task.
- Tools. What the agent is allowed to use: reading a repository, searching documents, querying a model of your architecture.
- Context. The material it works from, such as your code, documentation and interview notes, and nothing else.
- Guardrails. Limits on what it can access and do, and checks on what it produces.
- Human checkpoints. Fixed points where a person reviews the work before anything moves on.
When we say “agents” on this site, we mean agents inside a harness like this.
What an engagement looks like
1. Frame the question
We agree what decision the work should support, what is in scope, and what material agents may see. This is a conversation, not a task for a model.
2. Read everything
Agents work through codebases across many repositories, architecture documents and interview notes. Work that used to take weeks of sampling now covers the full material in days.
3. Map the current state
Agents draft application inventories, integration and dependency maps, and a list of risks and open questions. The founder checks the drafts against the source and against what people in your organisation say.
4. Develop options
Agents help produce options, trade-offs and roadmap drafts. The founder decides which options are realistic and what to recommend, and why.
5. Deliver and keep current
You get the deliverables in a form you can maintain. Because the analysis is built on your own material, it can be rerun when the landscape changes.
Who does what
| Agents | The founder |
|---|---|
| Read code and documents at scale | Frames the question with you |
| Draft maps, inventories and summaries | Checks every finding against the source |
| Generate options and roadmap drafts | Decides what to recommend |
| Rerun the analysis when things change | Owns the result and presents it to you |
We run on agents ourselves
We use the same approach in our own company before we recommend it to anyone else.
Our back office. Receipts and invoices are read, classified and filed by agents. Card bills are reconciled against the filed receipts, and anything missing is flagged. A payment list with check-digit validation is produced on a schedule. A human approves every payment.
Our architecture assessments. We built agent tooling for AI conversations across multi-repo codebases. It lets us ask questions of a whole IT landscape, such as which systems write to a given table or which services depend on a legacy integration, and get answers traced back to the code. That is how an assessment gets to a reliable current-state picture quickly.
Your data and trust
- What agents see. Only the material you share for the engagement, agreed at the start. Access is limited to what the task needs.
- Your choice of AI provider. If your organisation has approved AI services, we work in those, in your accounts and under your agreements. Otherwise we use Claude (Anthropic), GitHub Copilot and Microsoft 365 Copilot (Microsoft), and occasionally ChatGPT (OpenAI). Some of these process data in the United States, under EU-approved transfer terms.
- No training on your data. We use these services only under business terms or with settings where client data is not used to train models.
- Reviewed by a person. Every deliverable is reviewed by the founder before you see it.
- Deleted afterwards. Your material is returned or deleted within 90 days of the engagement ending, unless the contract says otherwise. See the privacy policy.
See how it would work for you
In 30 minutes we can go through a decision you are facing and how this way of working would apply to it.