The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
Why most "AI agents" fail in production: The difference between fragile scripts and deterministic engineering harnesses.
Over the past year, engineering teams rushed to integrate autonomous LLM agents into their workflows: automated PR reviewers, code migration scripts, internal chatbots, and automated incident responders.
Yet, within weeks, most teams abandon them: "The agent hallucinates non-existent CLI flags, gets stuck in infinite loops, breaks JSON schemas, and requires more human hand-holding than doing the work manually."
The root failure isn't the underlying model—it's the absence of systems engineering discipline around the agent harness:
1. Type-Safe Tool Calling via MCP: Ditch loose string-parsing scripts. Implement standardized Model Context Protocol (MCP) servers with strict JSON schema validation. When tools enforce strict type contracts, agents cannot hallucinate invalid parameters.
2. Role-Separated Multi-Agent Pipelines: Never let a single prompt write code, audit itself, and commit to production. Enforce separation of responsibilities: an Architect proposes the plan, a Coder implements, an Adversarial Reviewer attacks edge cases, and an Auditor verifies test-to-spec traceability.
3. Sandboxed Execution Enclaves: Run agent tool execution inside unprivileged, SELinux-confined container environments with network egress proxying to prevent prompt injection data exfiltration.
4. Self-Healing Test Harnesses: Build deterministic feedback loops where test failures are automatically captured, translated into reproducible test fixtures, and verified before any code touches main branch.
Stop treating AI agents as magical black boxes. Engineer them with the same deterministic rigor as your core infrastructure.
Architect deterministic agent systems with our 2–4 week Autonomous AI & Agent Workflow Integration Sprint on Contra: https://contra.com/s/POXsvYUQ-autonomous-ai-and-agent-workflow-integration
#ArtificialIntelligence #AIAgents #MCP #LLM #SoftwareEngineering #Automation #DevOps #Python
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Related posts
Behind the scenes on this one, because the build was more interesting than the result.
The task: get an AI agent to play a browser Tetris clone and beat the high score.
Problem one — the game renders to a <canvas>. No DOM, no list of filled cells, just pixels. So the first thing to build was a reader: divide the canvas into a 10×18 grid, sample each cell, decide filled or empty.
First attempt classified by colour, matching each cell against known block colours. Failed instantly and silently. Purple pieces read as empty. Gradient regions read as empty. The bot was playing against a board that was largely imaginary.
The fix was changing the question. Not "what colour is this cell" but "how much does this cell vary." Empty cells are flat. Filled cells have an edge, a bevel, a highlight. Measure luminance across the cell, take max minus min, threshold it. Works on every piece colour without enumerating any of them.
Then a flood fill to isolate the falling piece, and a placement search — for every column and rotation, simulate the drop, score the resulting board, pick the best.
Problem two — the browser protocol caps each execution at 45 seconds. Tetris doesn't finish in 45 seconds. So instead of an async loop that waits, you install an interval, return immediately, and poll the page every twenty seconds.
That's where it went wrong. The smart bot was too slow, so it got simplified down to a fixed key pattern. And the number being polled turned out to be the loop's own tick count, not the game score. Five runs of "improving" numbers that measured nothing at all.
The interesting part was the cleanup. To find out what actually happened, I went back to a 30-minute screen recording:
— ffmpeg to pull frames at 2fps
— locate the playfield in pixel coordinates
— classify every cell across a dense frame sequence
— then render the classifier's verdicts back over the frames and check them by eye
That last step is the one that mattered. The first pass flagged the bottom row as complete in almost every frame, which looked like a bug in my grid alignment. It wasn't. The row genuinely was complete, and the game genuinely never cleared it.
Verified: a completed line sits in the well for ten full seconds while the stack piles on top of it. Across ninety seconds of play the board empties about thirty times, and every one of those is a game over, not a line clear.
The lesson I'm taking into the next build: before you simplify under pressure, write down what you're optimising for, then check whether the simplification is still attached to it.
infiniteup.dev
The Tetris Clone Didn’t Clear Lines. Our AI Agent Didn’t Notice. - InfiniteUp
We pointed an AI agent at a browser game and it reported five runs of improving scores that measured nothing at all. It also reached a correct conclusion on worthless evidence. What that failure means for anyone deploying agentic AI.
tetris on a canvas is a tough one. did the agent beat the high score?
Automated a manual Google Sheets → Slack workflow to eliminate repetitive data entry and give the team instant notifications.
Previously, new data had to be manually checked, processed, and shared across the team — slow and easy to miss. I built a Make.com scenario that:
Watches Google Sheets for new/updated rows
Processes and routes the data through a router based on conditions
Creates corresponding Trello cards for task tracking
Logs processed records back to Google Sheets
Sends real-time Slack notifications so the team is instantly updated
Result: What used to be manual checking and updating is now fully automated — saving hours weekly and removing room for human error.
Tools: Make.com, Google Sheets, Trello, Slack
wow love this
Built an AI agent that handles access requests end-to-end inside Slack — reviewing, approving, and notifying automatically.
Manually tracking who requested access to what (and following up over email) was slow and easy to lose track of. I built a Make.com scenario that:
Watches Slack for new access request messages
Shows a typing indicator while the AI Agent processes the request
The AI Agent (Access Request Agent) evaluates the request and decides the next action
Depending on the decision, it logs the request in Google Sheets, sends an email via Gmail, or replies in Slack
Sends a confirmation message back to the requester in Slack
Result: Access requests are now reviewed and routed automatically — no manual follow-up, and every request is logged and answered consistently.
Tools: Slack, Make AI Agents, Google Sheets, Gmail (built on Make.com)
