Custom Telegram AI Agent Replaces OpenClaw for News Automation
Built CC Claw, a multi-CLI AI agent controlled via Telegram, with memory, evolution, and skills that automates news curation from scan to multi-platform posts—taking a month of iteration for stability over OpenClaw's limits.
Ditched OpenClaw for Full Control: The Custom Build Decision
OpenClaw handled basic news scanning for the author's Telegram channel (@GenAI_Spotlight), pulling from GitHub, Reddit, major media, and X accounts every few hours. But it lacked depth—no seamless backend switching, account rotation, or persistent memory across sessions. The author stripped OpenClaw to just scanning duties and built CC Claw (Coding CLI Claw) from scratch using CLIs for Claude, Gemini, CodeX, and Cursor as backends. Why? OpenClaw's one-size-fits-all couldn't evolve or handle nuances like process spawning/closing without losing context. Tradeoff: A month of nightly tuning (not hours as hyped videos claim), multiple refactors via spec → plan → review cycles, and ongoing bugs delaying release. Result: An agent that does everything OpenClaw did, plus gated permissions, verbose debugging, and Telegram-only interface for focus—no Discord/WhatsApp distractions.
'Oh, you can build an agent very easily, you know, like takes you few hours. I've seen some videos like that. No, if you really want to build a serious agent with memory and evolution... it kind of became a massive project.' (Author contrasts hype with reality; this sets expectations for serious builds requiring iteration.)
Multi-Backend Architecture with Persistent State
CC Claw spawns CLI processes per interaction, resuming with summarized context to mimic continuity. Switch backends (e.g., Gemini → Claude) via commands; it saves chats, offers 'don't save silly conversation' option, and uses a dedicated summarizer (Gemini Pro high by default). Account rotation for Gemini handles rate limits across personal API keys, auto-switching on timeouts—fixed OpenClaw's failure here. Permissions tiers: 'safe' (read-only tools), 'plan' (analyze/propose), 'gated' (manual approve/reject actions like file creation). Verbosity toggles show tool calls, model signatures (e.g., 'Gemini 1.5 Pro, high, key #2'), and actions. Shell access (/pwd, /restart) bypasses agent for direct backend control; /status reveals issues instantly, unlike OpenClaw's delays.
Skills system loads modular prompts (built-in, CC Claw-specific, universal manager downloads/scans for risks/prompt injection). MCPs (e.g., Perplexity for fact-checks, Google Search, NotebookLM) extend capabilities. Memory: Episodic with decay, browsable (/memory), slash commands to pin facts. Whiteboard logs session state (drafts, images) across backend switches. Chrome jobs run nightly: news scanner, GitHub PR alerts, YouTube stats, reflection (analyzes interactions for evolution suggestions), heartbeat.
Tradeoffs: CLI spawning adds latency/resume complexity; non-dev author spent hours debugging stability. But it enables 'yolo' mode for power users vs. sandbox hype.
'I don't believe in that approach sandboxed agents. And I know there are a lot of risks with letting your agent go yolo. You only live once.' (Author rejects locked-down agents; prioritizes utility with controls like gating.)
Newsroom Workflow: From Scan to Multi-Platform Distribution
Chrome job feeds proposals to Telegram 'forum topics' (newsroom chat). Author pastes URL (e.g., 'Agentic will redefine work'); agent invokes newsroom skill: dedupe vs. channel history, load voice/context (bold headline, spaced lines, 'why it matters'), generate image prompt, draft in HTML/Markdown for Telegram. Posts to staging channel for review; uses whiteboard for artifacts (draft path, image link). Inline edits fix issues (e.g., wordy → punchy, Flash model's formatting fails). Fact-check via Perplexity MCP: 'All core claims verified.' Author approves → agent posts to main channel (@GenAI_Spotlight), adds emojis, cleans whiteboard.
Then pushes to Buffer (API/MCP integrated): Adapts for LinkedIn/X (strips hyperlinks, appends channel promo: 'See more @GenAI_Spotlight'), queues/publishes/drafts per predefined slots. Python scripts enforce formats—no agent forgetfulness. Full cycle: 5-10 mins hands-on (review/edit/approve), automates scanning/curation/posting/distribution.
Before: Manual OpenClaw tweaks. After: End-to-end, multi-model resilient (switch on limits), human-in-loop for quality. No metrics shared, but daily posts shown (e.g., recent Agentic story live on Telegram/X/LinkedIn with matching image).
Agent Evolution and Optimization: Controlled Improvement
Nightly reflection scans interactions, flags issues (e.g., 'u re actions to review'), suggests identity/skills/context edits. Manual apply/reject/review/discuss—rejects auto-evolution to stay in control. Optimize audits: Souls.md (agent identity), identity.md, user.md for contradictions/misplacements; skills for token bloat/ineffectiveness. Universal skills manager imports MCPs securely.
Tradeoffs: Time-intensive (author skips auto for directionality), but yields tailored agent. Still: Context resume inaccuracies, Flash inconsistencies, gated timeouts.
'I don't believe in automatic evolution because what if the direction the AI is taking is not what I am interested in. So I like to be in control still not fully autonomous but that's my route.' (Author on evolution; emphasizes human oversight for alignment.)
'Everything that you see here, all these news, they're essentially being curated by the agent... We continuously scan multiple sources... and then I get a report with top stories that are aligned to my taste.' (Describes newsroom output; shows personalization payoff.)
Key Takeaways
- Start agent builds with spec/plan/review cycles; expect weeks for production stability, not hours.
- Use CLI backends + summarization for cheap multi-model switching with memory.
- Build whiteboards for state persistence across sessions/tools.
- Gate high-risk actions but enable 'yolo' for real utility—balance beats sandbox hype.
- Automate distribution via APIs (Buffer) with platform-specific scripts.
- Run nightly reflections for evolution, but manual review to control direction.
- Telegram-only UI simplifies; add /status/shell for ops visibility.
- Skills + MCPs + context files modularize workflows without token overload.
- Fact-check every draft (Perplexity MCP) before human approve.
- Test rigorously: Author iterated refactors after bugs in usage rotation, context.