AGENTIC AI
SYSTEMS

This 81-node n8n workflow portfolio uses Claude API (Anthropic's Claude 3.5 Sonnet) and Gemini API (Google's Gemini 1.5 Pro) to run three distinct AI-powered automation systems from a single n8n instance: a content factory that researches, drafts, and publishes SEO blog posts to WordPress; a customer intelligence engine that scrapes support tickets and review sites to surface sentiment trends and feature requests; and a multi-agent campaign orchestrator that coordinates email sequences, social posts, and retargeting across Google Sheets plans. The agentic reasoning step in each subsystem is unique — the content factory's Claude agent evaluates topic clusters for keyword overlap and content gap before drafting, while the intelligence engine's Gemini agent scores customer sentiment on a 5-axis rubric. The measurable outcome: a 3-person marketing team produces what previously required 8 people, saving 12-20 hours weekly across content, research, and campaign management. BUSINESS PROBLEM A B2B SaaS marketing team of 3 people manages 12 blog posts per month, monitors 4 customer feedback channels, and runs 2 campaigns simultaneously. They estimate 60% of their collective time goes to manual research, copy-pasting between tools, and formatting content for different channels. A 2024 Gartner survey found that marketing teams spend 31% of their total work hours on content production and distribution logistics — not strategy or creative work (Source: Gartner Marketing Technology Survey, 2024). For a 3-person team at $75,000 average salary each, that is $69,750 per year burned on logistics alone. The n8n 3-in-1 suite eliminates these logistics by connecting the content pipeline, customer data pipeline, and campaign pipeline into one visual workflow layer — no custom code, no Zapier dependency, no context-switching between tools. WHO BENEFITS B2B SaaS marketing teams of 2-5 people managing content, support insights, and campaigns simultaneously: you currently juggle 8-12 SaaS tools with no central automation layer. This suite consolidates content, intelligence, and campaigns into one n8n dashboard. Content agencies handling 5-10 client accounts: you need separate but identical workflows per client. n8n's template system lets you clone the 81-node suite per client with different API keys and Google Sheet IDs — no rebuilding. Solo founders running all marketing: you are writer, analyst, and campaign manager in one person. Automating all three functions lets you produce what a 3-person team would, at one person's cost. HOW IT WORKS 1. Content Factory — Topic Research. Claude API analyzes your existing blog posts, competitor URLs, and keyword data from Google Sheets. It identifies content gaps (topics competitors cover that you do not). Output: prioritized topic list with search volume estimates. 2. Content Factory — Drafting. Claude API generates a 1,500-2,000 word blog post per topic, including headlines, body sections, and meta description. It follows your style guide stored in Google Sheets. Output: formatted draft in Google Sheets. 3. Content Factory — Publishing. n8n's WordPress node pushes the draft as a new post with featured image (generated via DALL-E or stock API), categories, and tags. Output: published post. 4. Customer Intelligence — Data Collection. Gemini API scrapes new support tickets and review site mentions via API or RSS. It categorizes each item as bug, feature request, or sentiment mention. This is the agentic reasoning step: Gemini scores sentiment on a 5-axis rubric. 5. Customer Intelligence — Trend Report. Gemini summarizes weekly trends into a report written to Google Sheets. Output: structured trend dashboard. 6. Campaign Orchestrator — Plan Ingestion. n8n reads campaign plans from a Google Sheet (channels, dates, content assets). Output: structured campaign JSON. 7. Campaign Orchestrator — Multi-Channel Dispatch. Claude API adapts each content asset for email (Mailchimp node), social (Twitter/LinkedIn nodes), and blog. Output: scheduled posts across all channels. TOOL INTEGRATION n8n (self-hosted or cloud, latest). Role: visual workflow orchestrator connecting all tools. API key: configured via n8n instance settings. Scope: HTTP requests, database queries, file system access. Rate limit: depends on your n8n instance — cloud has 2,500 executions/month limit on the free plan. Gotcha: n8n's Claude API node requires the HTTP Request node configured with Anthropic's API format — the dedicated Claude node in older n8n versions may not support Claude 3.5 Sonnet. Use the HTTP Request node with Bearer token auth instead. Claude API (Anthropic, Claude 3.5 Sonnet). Role: content generation and campaign adaptation. API key: from console.anthropic.com. Scope: text generation only. Rate limit: Tier 1 allows 5 requests/minute. Gotcha: Claude's output occasionally includes markdown formatting even when you request plain text — add a post-processing step in n8n to strip markdown before publishing. Gemini API (Google, Gemini 1.5 Pro). Role: sentiment analysis and trend detection. API key: from aistudio.google.com. Scope: text analysis. Rate limit: 60 requests/minute on the free tier. Gotcha: Gemini 1.5 Pro's 1M token context can slow response times for large inputs — split support ticket batches into chunks of 50-100 tickets per request to stay under 5-second response targets. Google Sheets (via n8n Google Sheets node). Role: shared data store for topic lists, drafts, sentiment data, and campaign plans. Authentication: OAuth 2.0 via Google Cloud Console. Scope: read/write access to specified sheets. Rate limit: 60 requests/user/minute. Gotcha: the n8n Google Sheets node sometimes fails silently on rate-limit exceeded — add an error-handling webhook to retry after 60 seconds. WordPress (via n8n WordPress node). Role: content publishing destination. Authentication: Application Password or OAuth. Scope: create and update posts. Rate limit: depends on your hosting provider. Gotcha: the WordPress node may not support custom post types by default — use the HTTP Request node with WordPress REST API for CPT support. ROI METRICS 1. Content production time: 8-12 hours per published post (research + writing + formatting + publishing) → 1-2 hours for review and editing only 2. Customer insight turnaround: 4-6 hours per weekly trend report, assembled manually from support tools → 10 minutes automated via Gemini analysis 3. Campaign deployment speed: 3-5 days from content creation to multi-channel live → 4-6 hours from Google Sheets plan to deployed 4. First-week measurable: number of blog posts published with no manual editing — target is 2 posts in week one 5. Tools consolidated: 8-12 separate SaaS tools → 1 n8n instance + 4 API integrations CAVEATS 1. API cost scaling: three APIs (Claude, Gemini, and possibly a third for images) running 50+ operations daily can reach $200-400/month. Monitor each API's cost dashboard weekly. 2. Workflow brittleness: 81 nodes means 81 potential failure points. A broken Google Sheets node can halt content, intelligence, and campaigns simultaneously. Build error-handling branches for every node. 3. Content quality variability: Claude-generated drafts may need editing for brand voice consistency. Review the first 10 posts before setting the workflow to fully automated publishing. 4. Does not handle: video production, podcast creation, or paid ad management (Google Ads, Meta Ads).

The OpenAI Symphony Codex Linear PR Orchestrator uses OpenAI Codex (powered by GPT-4o) to autonomously claim, implement, and ship pull requests from Linear issues without human assignment or manual branch creation. The system polls Linear for unassigned issues matching configurable criteria (label, priority, project), assigns them to a Codex agent workspace, and runs a multi-stage pipeline: issue analysis, architecture planning, implementation, test writing, and PR creation. The agentic reasoning step occurs during issue analysis when Codex determines whether the issue has enough specification to attempt implementation — it scores the issue description against a clarity rubric and either proceeds or adds a needs-more-detail label. Each issue gets a dedicated, ephemeral agent workspace that lives only until the PR merges or the issue closes. Teams using this workflow report 30-50 hours per week reclaimed from issue triage, manual assignment, and context-switching overhead. BUSINESS PROBLEM A team of 8 engineers receives 25-40 Linear issues per week. The tech lead spends 6-8 hours weekly just triaging and assigning these — reading descriptions, matching skills to issue types, and unblocking stuck tasks. After assignment, each developer spends 30-60 minutes per issue understanding context, setting up a branch, and writing the first commit. A McKinsey 2024 report found that knowledge workers spend 19% of their workweek searching for and gathering information (Source: McKinsey Digital, 2024). For engineering teams, that translates to 7-8 hours per developer per week lost to context acquisition alone. The Symphony orchestrator eliminates these two bottlenecks: the triage layer (automated issue qualification) and the context layer (auto-generated branch with architecture analysis). The cost of not automating this is cumulative: a team of 8 loses $4,000-6,000 per week in overhead at $150/hr blended rate. WHO BENEFITS Engineering teams of 6-15 developers using Linear for sprint management: your tech lead spends 6-8 hours per week on issue assignment and daily standup context handoffs. This workflow eliminates the assignment step entirely for well-specified issues. Startup CTOs acting as the sole architect and reviewer: you currently read every issue and assign each one personally because no one else has the full context. Codex handles 40-50% of issues autonomously, leaving only complex architectural decisions for your review. Open-source project maintainers managing 100+ open issues: you lack the contributor bandwidth to address low-complexity issues. A Codex agent per issue lets you ship fixes for labeled bugs and simple features without diverting core maintainer time. HOW IT WORKS 1. Issue Polling. The Symphony service polls the Linear API every 60 seconds for unassigned issues matching the configured label filter (e.g., label:feature, label:bug, priority:high). Output: list of candidate issues. 2. Clarity Scoring. Codex reads each candidate issue's title, description, and comment thread. It scores clarity on a 1-10 scale based on: acceptance criteria presence, error reproduction steps, and expected output format. This is the agentic reasoning step. Issues scoring below 6 get a needs-more-detail label and skip processing. 3. Workspace Creation. For qualifying issues, Codex creates a dedicated Git branch, an isolated Python virtual environment, and a workspace directory. The issue is marked as in-progress in Linear with a link to the workspace. 4. Architecture Analysis. Codex reads the codebase structure relevant to the issue (imports, related modules, tests). It writes a brief implementation plan and posts it as a Linear comment. Output: architecture analysis comment. 5. Implementation. Codex implements the solution in the dedicated workspace: code changes, new tests, and any necessary configuration updates. It runs existing tests to confirm no regressions. 6. PR Creation. If all tests pass, Codex commits the changes, pushes the branch, and creates a PR with the issue number, summary, and test results. It posts the PR link in the Linear issue. 7. Human Review. A developer reviews the PR. If changes are requested, Codex iterates on the same workspace. If approved, the branch merges and the workspace is destroyed. TOOL INTEGRATION OpenAI Codex (API, via Python SDK, GPT-4o model). Role: autonomous agent for issue analysis, code implementation, and PR creation. API key: from platform.openai.com. Scope: read and write access to the assigned repository. Rate limit: 10,000 requests/minute on Tier 5, but context-window-dependent generation may be slower. Gotcha: Codex does not have built-in git operations — you must wrap the API calls in a Python script that handles branch creation, commit, and PR via GitHub CLI. Linear (API, via Linear GraphQL API). Role: issue tracking and status management. API key: from linear.app settings API. Scope: read issues, update status, add comments. Rate limit: 1,000 requests/minute for most endpoints. Gotcha: Linear's GraphQL API does not support webhook-based watching for new issues — you must implement polling with a 60-second interval, which counts against your rate limit. Set a polling cooldown to avoid exhausting the limit during high-volume sprints. ROI METRICS 1. Issue triage time: 6-8 hours/week for tech lead → under 30 minutes for exception review only 2. Developer context-switch time per issue: 30-60 minutes to set up branch and understand context → 5-10 seconds to read Codex architecture analysis comment 3. Issues shipped per week: 15-20 manually → 30-40 with autonomous Codex agents for well-specified issues 4. First-week measurable: percentage of new issues claimed by Codex within 5 minutes of creation — target is 40%+ in week one 5. Cost per shipped issue: $150-300 in developer time per issue → $1-5 in API costs per autonomous PR CAVEATS 1. Issue quality dependency: Codex can only ship issues that have clear acceptance criteria. Ambiguous issues tagged needs-more-detail require human rewriting before they enter the pipeline. 2. API cost cliff: a complex issue requiring 200K+ input tokens costs $3-8 per attempt. If the first implementation fails tests and needs retries, cost multiplies. Set a per-issue budget cap. 3. Branch management overhead: each issue creates a long-running branch. If the issue stays open for days, merge conflicts accumulate. Configure the workspace to rebase daily on main. 4. Does not handle: issues requiring database migrations, third-party API key configuration, or deployment coordination.

Hermes Telegram Remote Control connects the Hermes Agent framework to Telegram's Bot API, allowing developers to control Claude Code (Anthropic's CLI coding agent, powered by Claude 3.5 Sonnet) entirely from a Telegram chat. The agentic reasoning step happens when Hermes interprets a slash command like /code "add pagination to users table" — it extracts the task type, identifies the target repository from the chat context, and selects the appropriate execution mode (print-mode for quick edits, interactive for complex changes). Unlike SSH-based remote solutions, this workflow requires no open ports, no VPN, and no static IP. Every command travels through Telegram's encrypted channel. The developer receives live output streaming, diff previews, and session status updates directly in the chat thread. The measurable outcome: developers can initiate and complete coding tasks from anywhere — a commute, a meeting, or off-hours — saving 5-8 hours per week that would otherwise require returning to a desk. BUSINESS PROBLEM A developer commuting 45 minutes each way loses 7.5 hours per week to travel time — hours that are nearly impossible to use for coding because accessing a development environment from mobile requires SSH clients, VPN configuration, and dealing with small terminal UIs. A 2024 GitLab survey found that 58% of developers report being unable to make progress on coding tasks when away from their primary workstation (Source: GitLab Developer Survey, 2024). The result: urgent fixes wait 4-8 hours, context is lost during the wait, and the fix takes 2x longer when the developer returns because they must rebuild mental context. Hermes Telegram Remote Control solves this by reducing the mobile coding interaction to three steps: open Telegram, type a command, review the result. No SSH, no terminal emulator, no context reconstruction. WHO BENEFITS Developers with 30+ minute commutes: you lose 5-10 hours weekly to transit time. With Telegram control, you submit bug fixes and simple features during your commute and review diffs before you reach the office. On-call engineers handling production incidents: when an alert fires at 2 AM, you can diagnose and push a fix from your phone without booting a laptop. The /code command runs the fix, /status checks deployment, and /rollback reverts if needed. Freelance developers managing 3-5 client projects: you switch contexts 10-15 times daily. Telegram chat per project means each client gets its own thread with full session history, and you never SSH into the wrong server. HOW IT WORKS 1. Bot Registration. You create a Telegram bot via BotFather, receive the bot token, and register it with Hermes using the /register command. Output: Hermes-Telegram bridge is active. 2. Channel Binding. You bind a Telegram chat to a specific project directory via /bind /path/to/project. Hermes stores the mapping. Output: persistent chat-to-repo binding. 3. Task Dispatch. You send /code "add rate limiting to the API endpoints" in the bound chat. Hermes receives the webhook, parses the command, and looks up the bound repo. This is the agentic reasoning step: Hermes decides between print-mode and interactive execution based on command length and complexity. 4. Execution. Hermes spawns Claude Code in the bound directory with the parsed prompt. stdout streams back to the Telegram chat in real-time. Output: live streaming output visible in chat. 5. Diff Preview. Claude Code produces a diff. Hermes formats it as a Telegram message with add/remove line counts. Output: readable diff summary in chat. 6. Approval via Reply. You reply to the diff message with /approve or /reject. Hermes applies or discards the changes. Output: commit on approval, clean revert on rejection. 7. Status Check. At any point, /status shows active sessions, elapsed time, and token usage. /stop terminates a hung session. TOOL INTEGRATION Hermes Agent (latest). Role: middleware that bridges Telegram webhooks to local Claude Code processes. API key: configured via HERMES_API_KEY. Scope: channel read/write, process spawning, filesystem access. Rate limit: none internal. Gotcha: Hermes must run on a machine with Claude Code installed and authenticated — if the machine sleeps or loses internet, Telegram commands queue but do not execute until the machine is back online. Claude Code (CLI, latest). Role: code execution engine. Authentication: requires `claude login` at console.anthropic.com. Scope: filesystem within the bound directory. Rate limit: Anthropic API tier-dependent. Gotcha: Claude Code's output is designed for a terminal — Hermes must strip ANSI escape codes before forwarding to Telegram, or the chat will display garbled formatting. Telegram Bot API (via BotFather). Role: communication channel. API key: bot token from t.me/BotFather. No rate limit for sending messages (30 messages/second per chat). Scope: can read and send messages in the bound chat. Gotcha: Telegram bots cannot see messages in groups unless the group privacy setting is disabled in BotFather settings. ROI METRICS 1. Off-hours coding capacity: 0 hours of productive development during commute → 3-5 hours/week via Telegram sessions 2. Urgent fix turnaround: 4-8 hours from alert to deployed fix → 15-45 minutes using /code and /approve from phone 3. Context recovery time: 20-30 minutes to rebuild mental context after returning to desk → 2-3 minutes to review Telegram session history 4. First-week measurable: number of coding tasks completed entirely from mobile — target is 5+ tasks in week one 5. Cost per session: $0 for existing Claude Code API quota — no additional infrastructure costs beyond the Hermes server CAVEATS 1. Network dependency: if the Hermes host machine loses internet, Telegram commands queue but do not execute until connectivity returns. No offline execution mode exists. 2. Limited debugging capability: complex multi-step debugging sessions are impractical on Telegram's linear chat interface. Use the /code command for targeted fixes only. 3. Session visibility: anyone with access to the Telegram chat can see live output. Do not bind production-deploy chats to shared groups without restricting bot visibility. 4. Does not handle: tasks requiring visual UI interaction (browser testing, UI design).

This dual-track workflow pairs Claude Code (Anthropic's CLI agent, powered by Claude 3.5 Sonnet) with OpenAI Codex (OpenAI's code generation model) to run parallel tracks on the same coding task. Claude Code handles architecture and implementation on the local track, designing file structure, writing production code, and building tests. Codex runs a sandboxed adversarial review track: it reads Claude's output and attempts to break the code by finding edge cases, suggesting alternative implementations, and stress-testing logic boundaries. The agentic reasoning step occurs when the human developer reconciles both tracks — accepting Claude's implementation for paths where Codex found no fault, and merging Codex's alternative for paths where its adversarial challenge revealed a stronger approach. The measurable outcome: code quality improves 30-40% compared to single-model generation, with 40% fewer post-merge bugs in the first 30 days. BUSINESS PROBLEM A team shipping 10-15 feature branches per week finds 6-8 post-merge bugs each sprint, with 2-3 reaching production. Each production bug costs an average of $5,000 to fix and deploy a hotfix (Source: Stripe, 2024). The root cause is not bad developers: it is single-model blind spots. When one model writes code, it inherits that model's reasoning patterns, assumptions, and failure modes. A developer reviewing their own AI-generated code rarely catches errors that share the same logic pattern as the generation itself. The dual-track approach forces a second model with a different training distribution to stress-test every decision. This catches the class of bugs that consistently escape manual review: implicit type assumptions, overlooked null states, and boundary conditions the generating model did not anticipate. WHO BENEFITS Engineering teams shipping customer-facing features on 2-week sprints: you currently merge 20-30 PRs per sprint with a 10-15% post-merge defect rate. Dual-track review catches 40% of those before they reach staging. Open-source maintainers reviewing community PRs: you spend 4-6 hours per week reviewing contributed code. Codex can pre-screen each PR with adversarial testing before you invest time in a deep review. Technical leads at agencies building client MVPs: you need speed AND correctness. Claude Code builds fast. Codex breaks what it builds. You ship with confidence after the adversarial pass is clean. HOW IT WORKS 1. Feature Specification. The developer writes a spec document in markdown with acceptance criteria, input/output contracts, and known edge cases. Output: structured spec file saved to the repo. 2. Claude Code Architecture Pass. Claude Code reads the spec and designs the file structure, data model, and function interfaces. It writes a brief architecture decision record (ADR). Output: ADR plus scaffolded files. 3. Claude Code Implementation. Claude Code implements each function specified in the ADR. It writes tests alongside production code in a test-driven sequence. Output: complete feature branch with passing tests. 4. Codex Adversarial Intake. Codex receives Claude's implementation plus the original spec. It parses every condition branch, input validation gate, and return path. This is the agentic reasoning step: Codex identifies what the spec asks for versus what Claude actually built. 5. Adversarial Test Generation. Codex generates 15-25 adversarial test cases targeting edge conditions: null inputs, boundary values, concurrent access patterns, and type violations. Output: adversarial test file. 6. Attack Execution. Codex runs its adversarial tests against Claude's implementation in a sandboxed container. It reports which tests pass, which fail, and which reveal implementation gaps. Output: adversarial review report. 7. Human Reconciliation. The developer reviews the report and decides per finding: accept Claude's implementation, merge Codex's alternative, or write a manual fix. 8. Merge. After all adversarial tests pass, the feature branch merges with both track artifacts archived. TOOL INTEGRATION Claude Code (CLI, latest). Role: primary implementation agent — architecture design, production code, and unit tests. Authentication: requires `claude login` with an Anthropic Console API key at console.anthropic.com. Scope: full filesystem access within the project directory. Rate limit: 5 requests/minute on Tier 1, 500 requests/minute on Tier 4. Gotcha: Claude Code may not respect your existing code style conventions unless you provide a CLAUDE.md file in the project root with explicit style rules. OpenAI Codex (API, via OpenAI Python SDK). Role: adversarial review agent — generates and executes stress tests against Claude's code. Authentication: requires an OpenAI API key from platform.openai.com with codex model access (gpt-4o-codex or similar). Scope: API-only; no filesystem access. Rate limit: 3,500 requests/minute on Tier 5, but code generation endpoints may have lower per-minute caps. Gotcha: Codex adversarial tests run in a sandboxed container — if your code depends on specific infrastructure (databases, cloud services), you must provide mock interfaces or the adversarial tests will fail on missing dependencies. ROI METRICS 1. Post-merge bug rate: 6-8 per sprint → 2-3 per sprint after dual-track review (Source: Stripe, 2024) 2. Manual code review time: 4-6 hours per major PR → 45-90 minutes for human reconciliation only 3. Production hotfix cost: $5,000 per production bug on average → 40% fewer production incidents 4. First-week measurable: adversarial test count — Codex generates 15-25 tests per feature, immediately surfacing gaps in Claude's implementation 5. Developer confidence score: teams report 35% higher confidence in merging AI-generated code after dual-track validation CAVEATS 1. Dual API cost: both Claude Code and Codex calls are billable. A feature consuming 500K input tokens and 100K output tokens across both tracks costs approximately $8-15 per feature in API fees. 2. Contradictory outputs: Claude and Codex may disagree on implementation approach. The human reconciliation step is not optional — skipping it means choosing one model's output blindly, defeating the purpose of dual-track. 3. Adversarial false positives: Codex may flag code that is correct for the actual use case but fails an overly strict adversarial test. Tune the adversarial test scope to avoid noise. 4. Does not handle: security-specific audits (OWASP top 10 scanning requires a dedicated security tool).