Embedded Files
# "I sat and reviewed sessions": A demonstration of total incompetence in automation
## Quote from the lecture
https://youtu.be/Nm3MsnngCJg?si=4QWKn2MKNIyEsEGg
Sheiko says:
> "I prepared a script that exports all sessions. And we literally made employees export all sessions in one or two days. And then **I sat and reviewed them**. Why did I review them? Two things needed to be extracted. First — the person's mistakes. Second — the agent's mistakes."
He says this with pride. As if it's an achievement. As if it's an expert approach.
This is **a demonstration of total incompetence in automation**.
---
## Why this is shameful
Let's break down what he does **manually**:
1. Exports sessions (okay, there's a script)
2. **Sits and watches each session** (here's where the shame begins)
3. Looks for the person's mistakes
4. Looks for the agent's mistakes
5. Writes feedback to the team
Steps 2–4 are **manual text analysis**. Pattern searching. Error classification. This is the **most automatable task** that exists.
And a person who spends 45 minutes talking about AI implementation, about agents, about process automation, **sits and manually reads logs**.
---
## This was solvable with cron even before the AI era
Even **without AI**, even in 2010, this was trivially automated:
```bash
#!/bin/bash
# cron: 0 2 * * * /opt/scripts/session_audit.sh
SESSIONS_DIR="/var/agent_sessions"
REPORT_DIR="/var/reports"
DATE=$(date +%Y-%m-%d)
# 1. Collect all sessions for the day
find "$SESSIONS_DIR" -name "*.json" -newer "$REPORT_DIR/last_run" > /tmp/sessions_list.txt
# 2. Parse: look for agent errors (timeout, tool_call_failed, context_overflow)
grep -l '"error"' $(cat /tmp/sessions_list.txt) > "$REPORT_DIR/$DATE_agent_errors.txt"
# 3. Parse: look for bad prompt patterns (too short, no criteria)
# ... a 20-line shell script
# 4. Generate report
echo "Agent errors: $(wc -l < $REPORT_DIR/$DATE_agent_errors.txt)" > "$REPORT_DIR/$DATE_summary.txt"
touch "$REPORT_DIR/last_run"
```
This is **20 lines of shell script**. This is **cron at 2 AM**. This is **a report in your inbox in the morning**.
No "I sat and reviewed". No "we made employees export". Just cron, grep, and a report.
---
## And now with AI this is trivially easy
Currently, in 2026, this is solved **even more simply**:
```
Every day at 02:00:
1. Script collects all sessions for the day
2. Feeds them to an AI auditor with the prompt:
"Analyze the sessions. Find:
- Person's mistakes (poorly formulated task, missing context)
- Agent's mistakes (insufficient tools, lost context)
- Patterns of recurring errors
Format: structured report."
3. AI auditor produces a report
4. Person reads the report for 5 minutes
```
Instead of **"I sat and reviewed"** — **5 minutes of reading a report in the morning**.
This isn't rocket science. This is **basic automation** that any DevOps engineer does every day. And a person who calls themselves an expert in AI implementation **didn't do this**.
---
## Why this demonstrates incompetence
Here's what this approach demonstrates:
**1. He doesn't know how to automate.**
A person with 10 years in development, who works with agents, who talks about AI-heavy domains — **didn't write a 20-line cron script**. He preferred to sit and manually read logs. This isn't an "expert approach". This is **a lack of basic automation skills**.
**2. He doesn't understand what scaling means.**
If he has 10 sessions a day — okay, you can sit and watch. If 100 — already impossible. If 1000 — physically unfeasible. The manual approach **doesn't scale**. And an expert who doesn't understand scaling isn't an expert.
**3. He doesn't use AI for what AI is built for.**
The entire lecture is about how AI automates routine tasks. And right there he **manually does the routine** that AI automates in seconds. This is **a contradiction**. It's like a chef who talks about kitchen food processors but cuts onions by hand.
---
## What to do right
**Step 1: Automatic collection.**
Cron or scheduled task collects all sessions for the period. No "we made employees export". This should happen automatically.
**Step 2: Automatic parsing.**
Shell script or Python script parses JSON logs. Finds errors, timeouts, context overflows, bad prompt patterns.
**Step 3: AI audit.**
Feed the parsed sessions to an AI auditor. It classifies errors, finds patterns, generates a report.
**Step 4: Person reads the report.**
5 minutes in the morning. Not "I sat and reviewed". But "I read the summary and made decisions".
**Step 5: Evolution.**
Based on the report, you adjust skills, prompts, tools. The **team does this themselves**, not "I wrote feedback".
---
## Conclusion
"I sat and reviewed sessions" is not an expert approach. It is **a demonstration of total incompetence in automation**.
Cron and shell scripts existed **decades before AI**. Any DevOps engineer knows how to automate log collection and parsing. And now, with AI, it's trivially simple: feed logs to an AI auditor, get a report, read for 5 minutes.
A person who talks about AI implementation but **manually reads logs** is not an expert. This is a person who doesn't know how to automate. And that's their problem. But don't present this as an "expert approach". It's shameful.
Page updated
Google Sites
Report abuse