Day 21 of 28 · OpenClaw Challenge
When Things Go Wrong
⏱ 7 min
📊 Intermediate
Let's be honest: agents aren't perfect. Tools fail. APIs go down. The agent misunderstands your intent. A workflow fires at the wrong time.
This is normal. Today you'll learn how to handle failures gracefully — so your agent recovers instead of crashing.
The three things that go wrong
1. A tool fails. The email API times out. The CRM is down for maintenance. The web search returns no results. The tool simply doesn't work.
2. The agent misunderstands. You said "send it to the team" but the agent sent it to the client team instead of the engineering team. It took the right action on the wrong target.
3. The output is wrong. The agent completed the task but the result is incorrect — a summary missed key points, a draft used the wrong tone, a calculation was off.
Knowledge Check
Which of these is the hardest failure to catch?
A
The agent completing the task successfully but with incorrect output
B
The internet going down
C
A tool failing with an error message
D
The agent saying "I don't know how to do that"
Tool failures are obvious — you get an error. But when the agent succeeds with wrong output, you might not catch it until the damage is done. That's why reviewing outputs (especially for high-stakes tasks) is critical.
How the agent handles tool failures
When a tool fails, the agent doesn't just give up. It uses the agent loop — going back to the Think step to consider alternatives:
Check if there are any open support tickets from Acme Corp.
Checking your support system...
⚠️ The support API timed out. Let me try a different approach...
Found them through your email instead. You have 3 open threads from Acme Corp:
1. "Login issues on mobile app" — opened 2 days ago, assigned to Jake
2. "API rate limiting question" — opened yesterday, unassigned
3. "Feature request: bulk export" — opened last week, low priority
I used email search as a fallback since the support API was down. Want me to try the API again later?
↻ Replay conversation
Knowledge Check
When the support API failed, what did the agent do?
A
It gave up and told you to try later
B
It crashed and needed a restart
C
It made up the support tickets
D
It looped back to Think, chose a different approach, and found the information through email instead
The agent used the agent loop — when Act failed, it went back to Think and picked an alternative tool (email search) to accomplish the same goal. That's the adaptive behavior the loop enables.
Human-in-the-loop for risky actions
For actions that are hard to undo — sending emails to clients, posting publicly, deleting data — your agent should ask before acting:
I've drafted a follow-up email to the Acme Corp team about the timeline delay. Before I send it:
**To**: Tom Williams (VP Engineering, Acme Corp)
**Subject**: Re: Phase 2 Timeline Update
The email addresses the delay directly, proposes a revised timeline, and offers a weekly check-in.
⚠️ This goes to an external client contact. Want me to:
1. Send as-is
2. Show you the full draft first
3. Cancel — don't send
Show me the full draft first.
Here's the complete draft:
---
Hi Tom,
Following up on our conversation about the Phase 2 timeline. After reviewing with the team, here's our revised plan:
— Phase 2 kickoff: April 7 (moved from March 15)
— Milestone 1: April 28
— Delivery: June 13
I'd like to propose a weekly 30-min sync to keep things tight. Would Thursdays at 2pm work for your team?
Happy to discuss further.
Best,
Alex
---
Send, edit, or cancel?
↻ Replay conversation
Knowledge Check
Why did the agent ask for confirmation before sending the client email?
A
Because client emails cost more to send
B
Because sending to an external client is high-stakes and hard to undo — the agent knows to check first
C
Because it always asks before every action
D
Because it wasn't sure about the email address
The agent recognizes that external client communications are high-stakes. A wrong email to a client can damage a relationship. So it pauses, shows you the draft, and lets you decide. Low-risk actions (sorting your inbox, creating internal notes) don't need this level of approval.
Reviewing and correcting mistakes
When you spot an error, tell the agent. It learns:
"That summary missed the budget discussion. Always include financial details in meeting summaries."
"Don't use that casual tone with clients. Use my professional persona."
"You scheduled a meeting during my focus time. Remember: 1-5pm is blocked."
Each correction gets stored in memory. The same mistake is less likely to happen again.
Week 3 — what you've mastered
Day 15: Custom instructions and personas — tailoring how your agent communicates
Day 16: Slack/Teams — monitoring team channels and filtering noise
Day 17: Automation — triggers, schedules, and proactive tasks
Day 18: Memory management — fixing the three failure modes
Day 19: Cost optimization — model tiering and the 63% savings
Day 20: Workflows — reusable multi-step automations
Day 21: Error handling — graceful failures and human-in-the-loop
Next week is the final stretch: self-hosting, security, custom tools, and a mini-project that brings it all together.
Final Check
What's the healthiest attitude toward agent errors?
A
Expect them, plan for them, and use corrections to improve the agent over time
B
Ignore them — the agent will figure it out eventually
C
They should never happen — if they do, the agent is broken
D
Disable all automations to prevent errors
Agents aren't perfect — and that's okay. The key is building in safeguards (human-in-the-loop for risky actions), monitoring outputs, and correcting mistakes so the agent learns. Each correction makes it better.
🛡️
Day 21 Complete — Week 3 Done!
"Errors happen. Handle them gracefully, and your agent gets smarter every time."
Tomorrow — Day 22
Build a Client Follow-Up System
Week 3 mini-project — automate client follow-ups using everything you've learned.