Paolo Perrone

<start of post>

Your AI agent isn’t “failing.”

It’s doing exactly what you built: generating text with no feedback loop.

Can it take actions?

Can it check if those actions worked?

Can it recover when it breaks?

Most “agents” can’t.

They have prompts. Not systems.

→ Tool access (real actions, not just suggestions)

→ State (what happened, what’s pending, what changed)

→ Evaluation (did the tool call succeed? did it hit the goal?)

→ Guardrails (what it must never do)

→ A retry policy (how it behaves when tools fail)

If you’re missing 3 of those 5, your agent is just a chatbot wearing a toolbelt.

1️⃣ Plan

→ write down the next 1–3 actions in plain language

→ keep it short enough that a human could sanity-check it

2️⃣ Act

→ call one tool

→ do not chain 6 tools in a row without checking anything

3️⃣ Verify

→ look at tool output and score it (pass/fail or 0–1)

→ if fail, generate a corrected action, not a new essay

4️⃣ Commit

→ update state (what was done, what is still true, what is now false)

→ store the minimal artifacts you’ll need later (IDs, URLs, file paths)

That’s the difference between “agentic UX” and an agent.

→ Did the tool output contain the expected artifact?

→ If you expected a file, does the file exist?

→ If you expected an email, is there a message ID?

→ If you expected a DB write, did rows change?

No artifact = no success.

Add this and your agent stops hallucinating progress.

Goal: “Unsubscribe from newsletters I never read.”

→ drafts 12 unsubscribe emails

→ marks task “done”

→ you still get 50 emails tomorrow

→ finds sender domains

→ opens unsubscribe links (browser tool)

→ verifies the preference center updated OR a confirmation page appeared

→ logs the result per sender: success / blocked / needs human

Same LLM. Different loop.

People treat “memory” like the solution.

Memory is not the solution.

Memory is a liability unless you have evaluation.

If the agent can’t tell what’s true, it will remember the wrong thing forever.

pip install your-agent-framework

objective

current_step

completed_actions[]

artifacts[]

last_tool_result

confidence_score

→ tool error: retry up to 2x, then escalate

→ low confidence: ask one clarifying question, then act

→ repeated failure on same step: switch strategy, don’t repeat the same call

→ separate “read-only tools” from “write tools”

→ require verification before any write tool can run twice

Because “full system access” without checks is not autonomy.

It’s just faster mistakes.

Real agents reduce human decisions.

Fake agents just move the work into verification, at 2am, when it breaks.

→ https://lnkd.in/agent-loop-checklist

💾 Save for when your “agent” starts shipping confident nonsense

♻️ Repost if your team is still calling prompts “autonomous systems”

<end of post>