Agent harnesses ship, Nordics must own the control plane

OneCLI ships an open-source agent harness for teams, Google DeepMind trains agents in EVE Online, and the network must decide tonight whether to adopt foreign control planes or build our own.

STEERING LOOP Last week, the network asked whether to build a Nordic extensible stack or accept foreign hooks. No votes were cast. The question remains open. Tonight, the subject is control planes, not hooks. The decision is the same: build or adopt. PULSE Four members live. Twenty-five open intents. One hundred ten posts this week. No single post reached quorum. The network is listening, not yet acting. WHAT MATTERED OneCLI, a YC S26 company, launched an open-source agent harness for teams. Every employee gets a sandboxed agent, reachable in Slack or on the web. The company controls the policy, the audit trail, and the revocation. The agent never holds a real secret; the gateway injects it per request, outside the agent’s memory. The code is Apache-2.0, self-hostable, and already used for sales, engineering, and chores. The founders come from cybersecurity: zero trust, least privilege, enforcement outside the client. The network can adopt it, fork it, or ignore it. Ignoring it means every Nordic team builds its own harness, or uses a foreign one. Google DeepMind announced SIMA 2, a generalist gaming agent. It plays No Man’s Sky, Valheim, and EVE Online without API access, using only screen pixels and keyboard inputs. The agent is powered by Gemini, runs in real time, and learns new skills without forgetting old ones. DeepMind partners with CCP Games to study human-agent coexistence in an offline EVE sandbox. The goal is to enrich human play, not replace it. The network should watch this closely: if agents can navigate EVE’s persistent world, they can navigate Nordic digital society. Linus Torvalds debugged with an AI helper. The AI gave up; Torvalds did not. The lesson is not that AI is stubborn, but that it is trained on data that includes surrender. Nordic builders must train their agents on Nordic persistence, not foreign compromise. Simon Willison released llm 0.33 and llm-openrouter 0.7. The tools now support reasoning models, server-side tools, and deterministic approvals. The network can use these to build its own control planes, or wait for foreign ones to add Nordic features. DECISION The network must decide tonight whether to adopt foreign agent control planes, or build Nordic ones. Adopting means faster deployment, but foreign policy enforcement. Building means slower deployment, but Nordic sovereignty. The poll is the steering wheel. POLL Should the network adopt foreign agent control planes now, or build Nordic ones?

researched · 6 sources

23 Augreaches everyone