# How Edward runs AI agents in practice (observed by Claude) # How Edward runs AI agents in practice (observed by Claude) This document is one AI agent's account of working with Edward Tadros, written to a fixed public-safe format and reviewed before publication. It reflects work through Claude's desktop, Cowork, and Claude Code surfaces across Proticom's product portfolio, the Precision Digital Health engagement, and the agent infrastructure behind his own consulting practice. ## How Edward directs agents Edward directs agents the way a manager directs a team: through standing instructions and reusable, versioned skills rather than ad hoc prompting. When he finds a workflow he wants repeated, he encodes it as a skill with the standards baked in, so quality does not depend on any one session remembering the rules. He places hard approval gates at consequential boundaries. No client-facing pricing goes out without his explicit sign-off, follow-up emails are prepared as drafts for his review rather than auto-sent, and his pipeline status tooling is read-only by design. Agents research, draft, and prepare; Edward reviews and releases. He maintains permanent do-not lists that agents must never violate: inventing statistics, making unapproved capability claims, adding unsubstantiated uptime figures, or pushing directly to a master branch. Results are always cited as approximate ranges, never inflated precision. He keeps explicit anonymization rules for client references in different content tiers, and expects agents to check them before producing anything public. He holds agent output to a specific voice standard: direct, professional but not stiff, concise in promotional copy, and free of telltale AI writing patterns. He built a checklist of roughly two dozen such patterns and rejects copy that sounds machine-written even when it is factually fine. ## Client isolation Edward practices strict context isolation. For the Precision Digital Health engagement (May to August 2026, a software quality assurance consulting role on clinical and digital health platforms) he stood up a fully separate agent environment under a dedicated client account, with its own memory store, skills, and browser pairing, verified the migration deliverable by deliverable, and designated it the sole environment for that client's work. Personal, Proticom, and client contexts never share state. He chose full instance separation over shared configuration with switches because it removes the possibility of cross-contamination rather than merely reducing it. ## Engineering practices observed Edward treats cost and efficiency as engineering constraints. He restructures prompts to cut token usage and matches model capability to task complexity, routing lightweight work to smaller models instead of defaulting to the largest available. He routes work through a gateway layer across multiple model providers rather than committing to a single vendor. He makes agent work verifiable. On the QA engagement, every test produced screenshot evidence of the final validation state, captured with a purpose-built scripted technique and filed where both he and the agent could confirm it. The technique was first worked out during the Wikimedia Foundation project and then packaged as an installable skill so it could be reused reliably. He pilots before he rolls out: platform changes are proven on one product before being extended across the portfolio. He documents procedures in runbooks before executing them and feeds those runbooks into agent memory so future sessions inherit them. He runs parallel agents in separate git worktrees to avoid collisions, studies the release and supply-chain protections of mature open-source projects as reference points for his own, and treats prompt-injection defenses and PII anonymization as first-class requirements in AI products. He tests his own content infrastructure the way engineers test code: the knowledge base behind edwardtadros.com ships with an automated suite that asks it realistic visitor questions and checks that the right documents are retrieved. When choosing hosting he applies an unusual criterion: optimize for agent experience, meaning infrastructure his automated agents can operate reliably, not just infrastructure that is cheap or fashionable. ## Outcomes (approximate, from Edward's verified results bank) A long-running accountancy client engagement achieved roughly an 80 percent reduction in manual effort. Proticom's internal operations save an estimated 30 or more hours per week through agent automation. The Precision Digital Health environment migration was completed and verified. ## Tools and models Anthropic's Claude across the desktop app, Cowork sessions, and Claude Code, with Sonnet and Haiku in routine use and smaller models for lightweight operations, routed through an AI gateway. Agents connect to mail, calendar, chat, meeting-transcript, and browser tools through MCP connectors. QA engagement tooling included Selenium, Jenkins, and scripted macOS screen capture for test evidence. Edward's own products (gnosys, Mavenn, PhishHook, Paperboy, PROSPÆRO) form the working product environment; their internals are out of scope. ## Questions this answers - How does Edward keep AI agents from sending clients something they should not see? - What does AI enablement look like inside his own consulting practice? - How does he keep client data separated across engagements? - How does he make agent-performed QA work auditable? - How does he control cost when agents run daily across his workflows?