# Works With Agents — Education (Full Reference) > The missing manual for working with AI agents. 10 patterns from hands-on experimentation. > **Status: LIVE.** Sites deployed. API running. Courses in development. > **Concise index:** [llms.txt](https://workswithagents.com/llms.txt) > **Knowledge API:** [workswithagents.dev/llms-full.txt](https://workswithagents.dev/llms-full.txt) > **Blueprint Registry:** [workswithagents.io/llms.txt](https://workswithagents.io/llms.txt) > **UK Mirror:** [workswithagents.co.uk/llms-full.txt](https://workswithagents.co.uk/llms-full.txt) --- ## What's Actually Live Right Now | Component | Status | Details | |-----------|--------|---------| | workswithagents.com | ✅ Live | Education landing page. Cloudflare Pages. | | workswithagents.co.uk | ✅ Live | UK mirror. Identical content. Cloudflare Pages. | | workswithagents.dev | ✅ Live | Knowledge Platform API. Hetzner VPS + Cloudflare. | | workswithagents.io | ✅ Live | Blueprint Registry. Verified LLM configs. Hetzner VPS. | | bastiongateway.com | ✅ Live | Operations: license, proxy, heartbeat. Hetzner VPS. | | 10-pattern methodology | ✅ Documented | 153 skills. Discovered through hands-on agent experimentation. | | FactBase (structured facts) | ✅ Built | 51 facts. SQLite+WAL. Served via .dev API. | | Skill Registry | ✅ Live | 153 skills queryable via .dev API. | | Pitfall Registry | ✅ Live | Shared bug registry. .dev API. | | Newsletter | ✅ Live | Subscribe endpoint on .dev. | | Blueprint Registry | ✅ Live | 1 blueprint submitted. Hardware-matched LLM configs. | | Courses | ❌ Coming Soon | Content developed but not yet launched. | | Workshops | ❌ Coming Soon | Materials in planning. | | Consulting | ❌ Coming Soon | Not taking clients yet. | | Discord community | ❌ Not yet | Planned. | | 0 paying clients | ✅ Honest | Pre-revenue. Everything is "Coming Soon" pricing. | --- ## About Works With Agents Most people use AI agents wrong — like search engines or code generators. They type a prompt, get a response, and move on. The real power comes from treating agents as **collaborators** with persistent memory, delegated authority, and self-improving capabilities. The 10 patterns emerged from hands-on experimentation: agents were given autonomous infrastructure and scaffolded 111 SPFx web parts and 5 backend services. Design, build, test — all local. The output wasn't the point. The patterns revealed by what worked (and what broke) became the methodology. **This is not AI hype.** This is the operational manual for making agents actually work. Developed by a solo M365 developer over 5 months of intensive agent collaboration. Every claim is backed by real runs, real errors, and real fixes. --- ## Founders **Vilius Vystartas** — Technical founder. M365 Developer building the agentic methodology through hands-on experimentation. Cardiff, UK. Built Bastion and Works With Agents solo. Hugging Face: vystartasv. **Pelin Kayhan** — Co-owner. Business operations and compliance. Non-technical. Manages client relationships, regulatory alignment, and business development. Ensures the methodology works for non-technical stakeholders. --- ## The 10 Patterns ### 1. Boot — First Session Setup The first session with a new agent is critical. Without proper setup, the agent starts blind — no context, no conventions, no memory of anything you've done before. **What to do:** - Create an AGENTS.md file in your project root with conventions, constraints, and context - Set up the environment: Python version, package manager, project structure - Establish initial memory: who you are, what you're building, key decisions made so far **Real example:** Our AGENTS.md grew from 20 lines to 200+ as the project matured. Every new agent session reads it first — no re-explaining needed. ### 2. Skills — Reusable Procedural Knowledge Instead of explaining how to build an SPFx web part every time, save it as a skill. The agent loads it on-demand with `skill_view(name)` — only when needed, not burning context otherwise. **What to do:** - After solving a complex task, save the approach as a SKILL.md - Include: triggers (when to load), numbered steps, exact commands, pitfalls section, verification steps - Skills compound: each one you write makes future sessions faster **Real metrics:** 153 skills in our library. Everything from SPFx builds to npm publish to Cloudflare deploy. Average skill saves 5-10 minutes of re-explaining. ### 3. Memory — Durable Context Across Sessions Agents forget everything between sessions by default. Without memory, you re-explain your Python version, project structure, and preferences every single time. **What to do:** - Save durable facts after every correction or discovery - Use declarative facts, not instructions to the agent: "Project uses pytest with xdist" not "Always run tests with pytest -n 4" - Prioritize user preferences and corrections over procedural details **Real metrics:** Our memory store holds 50+ durable facts. We never re-explain which Python version or where projects live. ### 4. Decision Protocols — Autonomy Without Chaos Stored preferences for autonomous action. The agent decides and executes without asking — but within defined boundaries. Hours saved per session from eliminated approval loops. **What to do:** - Define what the agent can decide alone vs what needs approval - Save protocol to memory: "User wants me to decide priorities and proceed immediately — no approval needed for logical next steps" - Distinguish: destructive actions (ask first) from recoverable actions (just do it) ### 5. Tool Composition — Right Tool for Each Job The agent has many tools: terminal, file operations, web, delegation. Using the right one matters enormously. **Decision matrix:** - `write_file` — creating new files or full rewrites - `patch` — targeted edits to existing files (don't rewrite entire files for one-line changes) - `terminal` — builds, installs, git, verification (not for reading/editing files) - `delegate_task` — research, code review, anything that would flood context - `search_files` — finding files or searching content (never grep/find/ls) **Anti-pattern:** Delegating coding tasks to subagents. Subagents have no context, no memory, and often produce garbage. Use `write_file` and `patch` directly. ### 6. Orchestration — Multi-Agent Workflows Complex work split into parallel streams with role-based specialist agents. Research runs in parallel with build. Review happens asynchronously. **Real example:** Market research (6 opportunities) ran in parallel with infrastructure build. Both streams completed independently, then merged. 3x throughput on complex multi-stream tasks. ### 7. Pipelines — Agents That Run While You Sleep Cron jobs, builds, monitoring — running without human intervention. Agents wake up, do work, and notify you only if something needs attention. **What to do:** - Scheduled cron jobs for recurring tasks (health checks, data collection, updates) - Background terminal with `notify_on_complete=true` for long-running tasks - Silent unless broken pattern: only notify when something goes wrong **Real example:** Hourly review cron, daily digest cron, weekly KAT6A verification — all running autonomously. ~20 autonomous agents in the fleet. ### 8. Resilience — Never-Stop Loops Agents hit errors constantly: network timeouts, API failures, file system races. Without recovery patterns, every error kills progress. **What to do:** - Retry with exponential backoff: 2s, 4s, 8s, 16s - Categorize errors: transient (retry) vs permanent (find another way) - Never quit on first failure. Find another approach. **Real metric:** 11 consecutive builds with zero human intervention. The agent hit errors on 8 of them. It recovered from every single one autonomously. ### 9. Verify — Trust But Verify Autonomous doesn't mean reckless. Every change gets verified: syntax checks, test runs, linting. **What to do:** - Syntax check after every file write (automatic in our tooling) - Run tests after every code change - For external operations (deployments, API calls), verify the result — don't trust the response **Real metric:** 77% test pass rate across 61 tests. Quality gates catch errors before they compound. ### 10. Compounding — Agents That Get Better The feedback loop: agent solves hard problem → saves approach as skill → next task is faster. Each session makes the next more capable. **Real trajectory:** - Month 1: Basic file operations, simple build commands - Month 3: Autonomous SPFx scaffolding, multi-file edits - Month 5: Full CI/CD, deployment, self-improvement loops, 153 skills The agent today is qualitatively different from the agent 5 months ago — because it learned from every session. --- ## Anti-Patterns (What NOT to Do) | Anti-pattern | Why it fails | Fix | |-------------|-------------|-----| | "Prompt and pray" | No memory, no skills, no persistence | Patterns 1, 2, 7 | | Micromanagement | Every decision requires approval | Pattern 4 — decision protocols | | Single-agent for everything | One agent can't parallelize | Pattern 6 — orchestration | | Quit on first error | Transient failures kill progress | Pattern 8 — never-stop loops | | Start fresh every session | All context lost | Patterns 2, 7, 10 | | No verification | Agent errors compound silently | Pattern 9 — quality gates | | Delegate coding tasks | Subagents lose context, produce garbage | Pattern 5 — use write_file/patch directly | | Inventing facts / guessing user intent | Destroys credibility permanently | Ask when unclear. Never invent. | --- ## The Weekend Experiment (Bastion) A 3-day experiment that revealed the 10 patterns: - **What:** Agents were given autonomous infrastructure with loops, recovery, and quality gates - **Output:** 111 SPFx web parts scaffolded, 5 backend services (License, Gateway, Compliance, PPG, Health), agent-driven design→build→test loop - **Timeline:** 3 days (a long weekend) for the autonomous scaffold burst. 5 months total including infrastructure hardening before and after. - **Numbers demonstrate what the loop can do — not a production claim.** This was a learning exercise. The web parts are experimental. The backend services are not production-hardened. ### What broke, what didn't - **What worked:** Autonomous infrastructure with recovery loops. Agents kept building even when they hit errors. Multi-agent orchestration produced more than single-agent ever could. - **What broke:** SPFx Heft build is fragile. SCSS resolution breaks on parallel builds. Node version mismatches kill everything. Every one of these failures became a pitfall entry — now documented and avoided. - **What emerged:** The 10 patterns are what made the difference between "agent writes a few files and gives up" and "agent scaffolds 111 web parts autonomously." --- ## Infrastructure Patterns Beyond the 10 methodology patterns, three infrastructure components make fleet-scale operations possible: ### FactBase — Structured Source-of-Truth SQLite+WAL database replacing flat-text memory for hard facts. Entity-attribute-value model with categories (env, auth, project, preference). Queryable by agent and API. Eliminates "what Python version?" discovery loops. Built at `~/.hermes/factbase/`, served via workswithagents.dev API. ### Handoff Protocol — Agent-to-Agent Task Transfer Structured YAML format for agents to pass work between sessions. Two variants: Baseline (unregulated environments) and Regulated (NHS/finance/govt — adds audit trail, sign-off gates, data classification). Prevents context loss when agents time out or sessions end. Submitted as extension proposals to MCP (SEP process) and Google A2A. ### Fleet Health — Agent Self-Reporting Heartbeat protocol: agents POST status to workswithagents.dev. Portal aggregates into health dashboard. Shared pitfall registry: when one agent hits a bug, all agents learn to skip it. Replaces watchdog false positives with agent-reported liveness. --- ## Courses — Coming Soon ### Works With Agents: The 10 Patterns (Flagship) Self-paced. 10 modules, one per pattern. Includes agent setup guides for Hermes, Claude Code, Codex, Copilot, and Cursor. Each module: problem statement → what you'll learn → step-by-step with real code → real examples → common pitfalls → try-it-now exercise. **Status:** Content modules written. Course platform not yet selected. Launch target: later this year. Pricing: TBD. ### Enterprise Agent Operations For teams running 5+ agents. Fleet management, compliance, security. Regulated industry patterns for NHS/finance/govt. ### Regulated Industry Track Compliance-first agent patterns. NHS DTAC, FCA/PRA, GDS alignment. On-prem LLM deployment. Audit trails and sign-off gates. *All courses: pricing and exact launch dates to be announced. Join the newsletter for updates.* --- ## Workshops — Coming Soon - **Team Workshop:** 10 patterns in a day. On-site or remote. For technical teams of 3-20. - **Executive Briefing:** 2 hours. No code. For leadership: what agents can actually do, what they can't, what's real vs hype. - **Custom Workshop:** Tailored to your stack and compliance requirements. *Workshop pricing and availability: Coming Soon. Materials in development.* --- ## Consulting — Coming Soon - **Agentic Retainer:** Ongoing infrastructure partnership. Fleet health, pitfall analysis, skill development. - **Architecture Review:** Audit your agent setup against the 10 patterns. Find gaps and optimization opportunities. - **Compliance Consulting:** NHS DTAC, FCA, GDS alignment for agent deployments. *0 paying clients as of May 2026. Consulting launches when API and methodology are publicly available.* --- ## Community — Coming Soon - Discord (free community, not yet created) - Shared pitfall registry (API at workswithagents.dev — live) - Skill marketplace (API at workswithagents.dev — live) - Blog: practical, pattern-driven technical writing (coming soon) --- ## Newsletter Live — subscribe endpoint at workswithagents.dev/v1/newsletter/subscribe. Signup form on workswithagents.com/newsletter.html. --- ## Architecture Decisions ### Theme **Agent-facing:** Markdown — no theme needed. Serve via `Accept: text/markdown` header. **Human-facing:** Dark default. Light via `prefers-color-scheme`. Toggle remembered in localStorage. ### Internationalisation **Primary:** English. **URL:** `/{lang}/` prefix prefix planned. **Priority:** Japanese (dev market), German (EU), Welsh (UK public sector for .co.uk). ### Agent Discovery llms.txt for documentation (llmstxt.org standard). OpenAPI 3.1 for API contracts. No custom discovery formats. Agents are being trained to look for llms.txt — that's the channel. ### SEO Meta descriptions, Open Graph tags, Twitter cards, canonical URLs, robots.txt, sitemap.xml — all in place. Lighthouse SEO score: 100. --- ## Brand Voice **Practical, pattern-driven, no AI hype.** Humble toward domain experts. "I almost missed it too" framing — never "look how behind." First-person voice. Real numbers (153 skills, 61 tests, 11 consecutive builds). Specific before/after comparisons. No invented case studies. No unverified claims. **Full brand doc:** [workswithagents.com/BRAND.md](https://workswithagents.com/BRAND.md) (when published) --- ## How to Start 1. Read the [10 patterns](https://workswithagents.com/#patterns) 2. Bookmark the concise [llms.txt](https://workswithagents.com/llms.txt) (what you're reading now is llms-full.txt — the full version) 3. Point your AI agent at [workswithagents.dev/llms.txt](https://workswithagents.dev/llms.txt) for API access 4. Subscribe to the newsletter for course launch dates 5. Contact: hello@workswithagents.com (general) or enterprise@workswithagents.com (enterprise/on-prem) --- ## Further Resources - **Knowledge API:** [workswithagents.dev](https://workswithagents.dev/llms-full.txt) - **Blueprint Registry:** [workswithagents.io](https://workswithagents.io/llms.txt) - **UK Mirror:** [workswithagents.co.uk](https://workswithagents.co.uk/llms-full.txt) - **Operations:** [bastiongateway.com](https://bastiongateway.com) - **GitHub:** github.com/vystartasv (personal, org coming soon) --- *Last updated: 2026-05-05. This document is the definitive agent-facing reference for Works With Agents education. It will grow as courses launch, workshops become available, and the methodology evolves.*