After 3 months running AI agents in production across 3 SaaS products, the author shares what worked (GitHub MCP, Postgres MCP, Playwright MCP) and what broke (long tasks, auth walls, cost blowups, multi-tool orchestration errors), with a monthly cost of ~$430.
About 6 things I tried worked well enough to keep in production, another 4 broke or hit hard limits I couldn't work around. Been running AI agents (Claude Opus 4.7 mostly, migrated to 4.8 last month, some GPT-5.5 for comparison) across 3 SaaS products for 3 months. Bigger picture: Gartner projects 40% of enterprise apps will embed AI agents by end of 2026 up from under 5% last year, and the MCP SDK hit 97M monthly downloads, so the adoption is real but production is messier than the demos. What worked. GitHub MCP for PR triage and code review saved roughly 8-10 hours a week, agent reads diffs, flags issues, drafts review comments I approve. Postgres MCP for read-only DB queries handled ~30 support tickets a week without me touching them, Claude writes the SQL, I approve, response goes out. Playwright MCP for QA on critical flows caught 4 regressions last month that would've shipped. Context7 for real-time docs stopped Claude hallucinating library APIs which alone paid for itself. Social scheduling via MCP shipped too. PostFast for cross-platform posting from Claude, 11 platforms including Google Business Profile, €10/mo, MCP works with Claude and ChatGPT. Metricool ($22/mo) handles analytics since PostFast's are thinner. Together they saved ~5 hrs/week on manual scheduling. Cons: PostFast community is small so docs on edge cases are lacking, Metricool has no n8n node so it only works if you drive it from Claude directly. What broke. Long-running agent tasks over 15 minutes stayed unreliable, they lose context or hit rate limits mid-flow. Anything with browser sessions behind auth walls (LinkedIn scraping, some SaaS logins) breaks constantly, Playwright can't hold session state well enough. Cost blowups on Claude Opus are real, my first month API bill hit $340 in one week when I let it run unsupervised on a research task. Fix was aggressive prompt caching (cuts cached input 90%) and defaulting research work to Sonnet 4.6 at $3/$15 per MTok instead of Opus 4.8 at $5/$25. Multi-tool orchestration across 5+ MCPs at once, agents pick the wrong tool maybe 20% of the time. TikTok posting via any MCP scheduler is still half-broken because of TikTok's API restrictions, PostFast, Blotato and Postiz all hit the same wall. Security is the part nobody talks about enough. Prompt injection is OWASP's #1 LLM vulnerability in 2026. A recent audit found 41% of public MCP servers have no auth at all and only 8.5% use OAuth, plus 30+ MCP-specific CVEs filed in a single 60-day window early this year. Stick with vendor-maintained servers (GitHub, Anthropic reference, official Metricool, official PostFast), don't just install random ones off Glama's 22K+ directory. Monthly cost after optimization: $200 Claude Max 5x plan + ~$150 API overflow on Sonnet + €10 PostFast + $22 Metricool + $50 hosting = around $430/mo. Cheaper than a part-time hire but you're still babysitting, so it's an assist not a replacement. Anthropic's own Claude Code numbers put typical devs at $150-$250/mo and heavy users at $500-$2000/mo, so my spend aligns with that band. What are you running in production successfully that I might be missing?? Especially interested in multi-tool agent orchestration wins since that's where I keep hitting the ceiling
A discussion on the challenges and successful strategies for deploying AI agents in production at scale, covering common pain points and effective solutions.
A discussion about deploying multi-agent AI systems in production, where different agents handle planning, execution, communication, and project management, asking about real-world experiences and bottlenecks.
A developer shares a personal experience of unexpectedly high costs from a multi-agent AI system, sparking a discussion on cost tracking and observability in agent frameworks.
The author describes a setup where different AI models are assigned to specific roles (planning, coding, review) to reduce API costs for a 24/7 autonomous engineering team, and shares common failure points like model wandering and hallucinated ownership.
Microsoft shares insights from shipping thousands of production AI agents at enterprise scale, covering the engineering challenges of moving from prototype to production, including the agent harness, retrieval-as-a-subagent, agent identity, and rubric-based evaluation loops.