In a recent deep dive into internal AI infrastructure, Bold.org engineer O. Ioannou detailed the evolution of an observability tool into a company-wide agent. What started in January 2026 as a read-only SQL interface for ClickHouse backend events now consumes billions of tokens monthly, serving engineering, support, product, and marketing teams. The retrospective offers a raw look at the friction points of scaling agent adoption, emphasizing that visibility and simple security boundaries mattered more than architectural elegance.

Visibility Drives Adoption

The most critical lesson was avoiding the 'build it and they will come' trap. Instead of launching a polished standalone web app, the team integrated the agent directly into existing workflows like Slack and Linear. By allowing real usage to occur in shared channels, the agent effectively marketed itself. Product managers and support staff saw colleagues solving real problems, which demonstrated capabilities far better than any documentation or launch email could. This organic discovery process turned every Slack interaction into a live demo of the agent's utility.

Security and Simplicity Wins

Ioannou adopted a deliberately conservative security model: read-only access gated behind Google sign-in for the company domain. By refusing to grant write permissions to remote systems, the team avoided the complex headache of ensuring the LLM wouldn't accidentally corrupt production data. The integration layer relied on boring, thin CLI wrappers around OpenAPI endpoints rather than waiting for mature MCP servers. This approach allowed the agent to aggregate context from PostHog, Grafana, and Slack without exposing credentials directly to the model, keeping the attack surface minimal while maximizing contextual awareness.

The Cost of Over-Engineering

The retrospective admits to significant time wasted on premature optimization. The team initially over-invested in complex evaluation frameworks and tracing infrastructure, trying to measure hypothetical failures. In reality, an internal tool does not need the rigorous eval pipeline of a frontier lab. The author also noted that engineers were actually harder to convert than non-technical staff, as they resisted abandoning their personalized local terminal workflows for a shared company interface. The biggest architectural miss was failing to design for local agent harnesses earlier, forcing users to leave their local context to access company data.

Key Takeaways

  • Integrate agents into existing communication tools (Slack/Linear) to drive organic adoption through visible usage.
  • Enforce strict read-only boundaries and simple authentication (SSO) to mitigate security risks.
  • Avoid premature complexity in evals; build infrastructure only to solve observed, recurring failures.
  • Recognize that non-engineers often adopt agent tools faster than engineers due to lower resistance to new interfaces.

The Bottom Line

Agent adoption is a social problem, not just a technical one. Stop building for architectural purity and start building for visibility; if people cannot see the agent solving their specific problems in their existing workflow, the most elegant architecture in the world will remain unused.