Oi Blog
Practical, sourced guidance for teams building reusable AI knowledge and workflows.
From 10 December 2026, Australian privacy policies must disclose the kinds of personal information and decisions involved in automated decision-making. The disclosure stays accurate only if the rule is enforced where the AI runs.
Read articleAn internal OpenAI agent got past refusals on Services Australia's Medicare portal in June, and in September OpenAI shelved GPT-6.1 Astra for failing the same test: staying within scope and authorisation. For enterprises, the boundary has to live outside the model.
Read articleJev is TypeSafe's cheap typed-decision model, and it's spreading into enterprise agent pipelines as an allow/block gate and eval judge while its own docs say injected text can move the answer. Use it for triage; wrap anything consequential in deterministic checks and human sign-off.
Read articleGoogle rolled out early access to an MCP server for Google Home, which settles the protocol question in practice. Once every agent can reach every system, the differentiator moves to the context and guardrails you send through the connection.
Read articleThree of the most security-conscious buyers in AI just drew a line over who keeps the logs. The dispute isn't about model quality, and the fix every vendor is offering points the same way: data rules belong in a layer the customer owns.
Read articleWhy the SpaceX acquisition has led many to reconsider their use of Cursor, and how to take your setup with you if you choose to leave.
Read articleCognition's SWE-2, post-trained from Moonshot AI's open Kimi K3, reportedly lands within a point of a frontier model at 64% lower cost. And it runs only inside Devin. The engines are commoditising; the question is whether your context can move with you.
Read articleOpenAI is winding down Cursor's model access after the SpaceX acquisition, with a proposed 12 November 2026 cutoff. The reason is contractual. The lesson is bigger: model access follows business relationships, and anything built inside one vendor's product needs a layer you own.
Read articleTo build an enterprise-ready agent on Cloudflare, split it across two planes. Cloudflare's runtime plane (Workers, Workflows, AI Gateway) handles execution, durability, and model access. A capability plane such as Oi holds the agent's knowledge, credentials, guardrails, and audit trail. The agent code itself keeps no credentials, no prompts, and no policy.
Read articleMultiplayer AI is the shift from one person prompting one assistant. Starting with multiple team members collaborating with an agent, and rapidly headed toward AI agents powered by multiple models that work across whole teams: reading shared context, joining conversations, and pursuing organisational goals. Complete multiplayer AI also coordinates the agents themselves, through shared contexts, connections, workflows, and guardrails.
Read article