Danilo Campos is a self-taught software developer, interaction designer, and PostHog robopsychologist who architected the PostHog AI Wizard, an autonomous coding agent that installs product analytics inside unfamiliar applications. He also created Agent Accords, a framework for giving people and AI systems shared software-development principles.
From mobile products to autonomous agents
Campos developed mobile applications for Hipmunk and Level Money during the early iPhone era. At Hipmunk, his attention to privacy shaped a travel-search feature that processed calendar information without sending it to company servers.
By 2016, he was GitHub’s technical director of social impact, contributing to ConnectHome’s technology-inclusion efforts. A self-taught programmer who spent part of his childhood in public housing, he helped bring computing instruction to public-housing residents. He later helped EveryoneOn replace costly, vendor-dependent infrastructure with existing services and custom software, documenting the transformation in his account of rebuilding the nonprofit’s systems.
His hardware projects include ThermTerm, an open-source heat-pump controller and home-automation terminal, and DeskHog, PostHog’s open-source developer device, for which he built initial provisioning, analytics, and companion screens.
At PostHog, Campos designed the AI Wizard to inspect existing applications, identify meaningful product events, and implement integrations appropriate to their architecture. His engineering account of the system describes its evolution from brittle single-shot generation into an agent capable of exploring repositories, consulting current documentation, and correcting mistakes.
Principles for dependable coding agents
Fresh context over stale model memory. Campos supplies current, framework-specific documentation to prevent invented APIs and outdated integration patterns; compiler output and type checking provide concrete feedback about generated changes.
Model Airplanes as architectural examples. Lightweight reference applications preserve recognizable patterns such as authentication and event tracking without the complexity of production software. Flattened into Markdown and included in agent skills, these Model Airplanes demonstrate what a successful integration should resemble.
Breadcrumbing through progressive disclosure. The agent first identifies business-critical code, then proposes useful analytics events, and only afterward modifies the application. This sequence limits architectural inconsistency while preserving flexibility in unfamiliar repositories.
Privacy boundaries and agent feedback. Constrained tools check or update approved environment keys without exposing entire configuration files. Post-run feedback also surfaces contradictory instructions, unavailable tools, missing permissions, and documentation targeting the wrong programming language.
Campos extends these ideas through Agent Accords, which packages reusable development guidance for humans and agents. His writing on product taste emphasizes familiar interface conventions, disciplined visual hierarchy, prototyping, and attention to details that earn users’ trust.
