AI and software engineering.
Teams that build with agents ship faster. Teams that do it without rules, metrics, or guardrails pile up debt, risk, and cost. The challenge is to industrialize.
Agentic development factory
What gets put in place:
- End-to-end agentic pipeline
- from ticket (Jira, Linear) to pull request—specification, TDD implementation, tests, documentation, PR—orchestrated in CI (GitHub Actions or equivalent).
- Adversarial agent review
- quality, security, performance, before any human review.
- Explicit human sign-off
- stricter review rules for agent-written code (dual approval, off-limits areas, what never ships to production without a human).
- Sandboxed agents
- isolated execution, controlled access and secrets, least privilege.
- Context engineering
- repository instructions (CLAUDE.md, AGENTS.md), conventions, architecture documentation agents can read, skills, connections to internal tools through MCP. The quality of the context determines the quality of the output.
- Tooling
- Claude Code and equivalent assistants, selected and configured for the team and its constraints.
- Metrics
- lead time, rework rate, automatic resolution rate, incidents, cost per task—before, during, and after.
- Team and roles
- training on the team’s own code, internal champions, new roles (specification, review, agent supervision). The bottleneck moves to product specification: it needs to be organized.
AI in production
AI features in the product.
- Evaluation
- reference datasets (golden sets), automated and human evaluations, regression testing on every prompt or model change, cross-validation across models from different vendors.
- Hardening
- specialized agents rather than one generalist agent, structured and validated outputs, a semantic layer over the data, fallbacks, a human in the loop on costly decisions.
- Observability
- traces, latency, error rate, cost per request, user feedback.
- Cost
- the right model for each task, caching, batch processing, context window management, a tracked budget.
- Vendor dependency
- abstraction layer, multiple models, exit plan.
Security and sensitive data
- LLM gateway
- (LiteLLM or equivalent): a single entry point to the models, prompt filtering, detection and blocking of personal data, logging.
- Hosting
- infrastructure that meets sector requirements (certified health data hosting, HDS, in healthcare), choice of regions and providers.
- Local open-weight models
- when the data must not leave.
- Traceability
- agent and human actions logged and auditable, as regulated sectors require.
- Prompt injection and exfiltration
- guardrails, tests, least privilege.
AI roadmap — 7 days
An assessment of how AI is used across the team and the product, followed by a prioritized plan to reach the next level: time saved, costs under control, quality maintained.
Then, if you decide to go ahead: implementation with the team.
7 working days over three weeks
Fixed fee — pricing on request
Let’s talk about your situation.
A few lines are enough: the company or portfolio company, the product, what’s happening, and what needs deciding.
Talk about your situation