Beyond the prompt: building an AI that takes initiative
In this GitHub Universe 2026 session, Cameron Hamilton, creator of Alliance AI, uses examples from HALIE to explain how agents can move beyond responding to instructions. He focuses on making that initiative dependable, measurable, and subject to human oversight.
From instructions to useful initiative
Hamilton describes two examples of HALIE taking a broader view of a request:
- When asked to connect a nursery camera, HALIE turned it into a baby monitor without an explicit instruction to do so.
- When asked to automate business growth, it suggested meeting customers in person rather than limiting its response to automation.
These examples frame the session’s central question: how can an agent decide what to do next while keeping its actions useful and accountable?
Three core architectures
Hamilton organizes the discussion around three architectures for proactive agents:
- Anticipation: deciding what to do next beyond the immediate prompt.
- Self-evolution: refining the agent’s own tools.
- Metacognition: examining the agent’s own decision-making.
The session connects these architectures to validating initiative and value, ensuring the system is wired correctly, and learning from GitHub.
Testing, cost, and human oversight
Hamilton’s practical framework emphasizes testing agents with real data and using measurable results to determine how much autonomy they earn. Human involvement remains part of the process rather than being removed as agents become more self-directing.
The session also addresses cost and performance, including running agents on cost-effective models alongside GitHub Copilot. The focus is on demonstrating useful results while controlling the expense of autonomous operation.
Session chapters
- 00:00 — Introduction to proactive AI agents
- 02:00 — Three lessons for building agents
- 05:26 — Core architectures of HALIE
- 07:01 — Validating AI initiative and value
- 09:49 — Ensuring proper system wiring
- 13:25 — Self-evolution and learning from GitHub
- 19:08 — Testing agents with real data
- 24:59 — Optimizing cost and performance
- 26:54 — Practical framework for building agents