Edit Summary
_w7D_hpnALZdmu_hIFODX5NOuPibermCgQOxzE5EChGF-kff_cxXPOdzDEWvPicyHg0C7GsldmNmNT8_W7Z6pwCrq5lYLa3YPNs4zJnEZrdXFXHORnbsFkrz_w==
Back
Edit the source note directly. Frontmatter stays at the top.
0. Properties - Meeting Date: 2026-03-06 - Meeting Type: 1x1 - Note Type: Summary - Attendees: Scott Warner; Shawn Remick 1. Meeting Summary We reviewed growing requests from technical teams to use terminal-based AI agents, specifically Claude Code, and contrasted that with our current approved tools. I reiterated the policy stance to use the tools we provide, while acknowledging the legitimate developer desire for console-first workflows that keep processing and files local. Claude business pricing is roughly $1,200 per year for five users, which is not terrible, but it risks license sprawl and inconsistent usage if we do not set guardrails. We aligned on enabling innovation without creating shadow IT. The right approach is to stand up a controlled dev sandbox environment for qualified technical users, with per-user VMs we can administer and observe, segmented on a dedicated Innovation Lab VLAN/subnet. Default outbound 443 would be allowed with managed firewall pinholes as needed. We need to correct the current Linux repo blocking in the lab so it is actually usable for work like Shaper development. I demoed my Codex-based terminal workflow that operates on local content and outputs structured artifacts. It generated OKR/KPI dictionaries, a readme, a monthly update schedule, and a baseline dashboard from meeting transcripts, which saved Vera significant time. Shawn raised a good question about whether Microsoft Copilot, scoped to a specific folder, could achieve similar outcomes; we need a head-to-head comparison. We also discussed the slow uptake of the AI training program and the need to move leaders from crawl to walk. The goal is to build capability in the functions, not to route hands-on work through me. Lastly, we noted a networking issue where VPN interferes with terminal tooling performance, which IT will address. 2. Attendee List - Scott Warner - Shawn Remick 3. Action Items - Scott Warner - Draft an AI tooling policy for technical users: approved tools list, up to two AI licenses per person from a defined set, request and justification workflow, data handling rules, and local-file usage expectations. - Scott Warner - Prepare a 15 minute demo of the Codex CLI workflow for IT and technical leads; document pros and cons, data boundaries, and when to prefer this over Copilot. - Scott Warner - Run a quick experiment to compare Copilot with a folder-scoped context versus the terminal agent workflow, and document whether Copilot is sufficient for OKR/KPI synthesis. - Shawn Remick - Design and stand up a controlled dev sandbox offering: per-user VM with local admin, logging/telemetry, and remote drive visibility, limited to an Innovation Lab VLAN/subnet. - Shawn Remick - Coordinate with Wolfram to implement the VLAN and firewall model for the Shaper Linux environment; ensure Linux repos are accessible in the lab while maintaining broader enterprise controls. - Shawn Remick - Fix VPN routing and firewall rules so terminal-based AI tools do not require disconnecting the VPN and perform reliably. - Shawn Remick - Follow up with Ronnie about any installs on Eric's machine; prevent shadow IT by moving interested users into the sandbox model. - Shawn Remick - Drive AI training enrollment with ELT and priority users; ensure Vera enrolls and begins hands-on practice. - Shawn Remick - Schedule a 45 minute decision meeting with IT and technical leads to finalize the AI tooling licensing approach and the lab model. 4. Relevant Timelines - Ronnie is out today and Monday; follow up on Eric install checks after Monday. (Shawn) - AI training has open seats now; enroll priority users asap. (Shawn) - Schedule the AI tooling policy decision meeting as soon as practical. (Shawn) 5. Additional Notes - Console-based AI agents offer real benefits for power users by keeping processing local, avoiding upload limits, and handling large project directories with fewer constraints. - Copilot can be folder-scoped; it may cover some scenarios adequately. A structured comparison will clarify when to use Copilot versus terminal agents. - We must avoid license sprawl and unmanaged installations. A sandboxed VM on an Innovation Lab VLAN strikes the balance between innovation and governance. - Current Linux repo blocking made the previous lab unusable; we need a clear, documented process for lab-specific firewall pinholes. - Broader education is needed to differentiate AI from automation and ML, and to set realistic expectations for outcomes. - Networking performance matters. VPN behavior currently degrades terminal agent performance and should be corrected.
Save