The Fix That Broke My Agent in a Different Way

This is part of a series on building a PR automation system. The previous post covered how my CICDAgent hallucinated a detailed CI failure report without calling a single tool — and how adding MANDATORY CONSTRAINTS to the prompt fixed it.

The MANDATORY CONSTRAINTS worked. The agent stopped hallucinating. I merged the PR and moved on.

Then I checked the next trace.

---

What happened this time

CICDAgent (agent)
├── get_workflow_runs (tool)
├── get_workflow_runs (tool)
├── get_workflow_runs (tool)
├── get_workflow_runs (tool)
├── get_workflow_runs (tool)
├── get_workflow_runs (tool)
├── get_workflow_runs (tool)
├── get_workflow_runs (tool)
├── get_workflow_runs (tool)
└── get_workflow_runs (tool)

11 spans. 10 identical tool calls, same input, same output, every single iteration. The agent never called get_run_logs. Never called analyze_failure. Just get_workflow_runs, over and over, until it hit the 10-iteration limit and gave up with: “I was unable to find an answer within the allowed number of steps.”

Signal	Value	Verdict
Span count	11 (1 agent + 10 tool)	Active but stuck
Unique tools called	1 of 4 needed	Looping
Iterations	10/10	Hit the limit
Output	”unable to find answer”	Failed

The hallucination was gone. A new failure had taken its place.

---

The root cause was in my own prompt

The constraint I’d added in the previous fix:

"MANDATORY CONSTRAINTS:\n"
"- You MUST call get_workflow_runs before any other tool or your Answer.\n"  # ← this line
"- You MUST call get_run_logs and analyze_failure before writing your Answer.\n"

That first constraint has two valid interpretations:

Intended meaning	What the LLM read
Call it once, at the start, before anything else	Call it before every other tool — always call it first at each step

The LLM followed the second interpretation. Each iteration played out like this:

Iteration 1:  call get_workflow_runs  ✓ (satisfies "before any other tool")
Iteration 2:  about to call get_run_logs...
              but constraint says: call get_workflow_runs before any other tool
              → call get_workflow_runs again
Iteration 3:  same logic → get_workflow_runs again
...
Iteration 10: same → MAX_ITERATIONS → give up

The tool returned the same result every time. The LLM was stuck in a rule it couldn’t escape, and never progressed past the first tool call.

---

The fix: three words

"MANDATORY CONSTRAINTS:\n"
"- Call get_workflow_runs exactly once as your first tool call, then move on.\n"  # ← changed
"- You MUST call get_run_logs and analyze_failure before writing your Answer.\n"

Three changes:

1. “exactly once” — eliminates the loop. One call, done, never again.
2. “as your first tool call” — scopes the timing to the start, not to every subsequent step.
3. “then move on” — explicitly tells the LLM what to do after. Closes the question of what follows.

After this change, the trace looked right:

CICDAgent (agent)
├── get_workflow_runs (tool)     ← called once
├── get_run_logs (tool)          ← moved on
└── analyze_failure (tool)       ← finished the job

---

The underlying lesson

Constraints create edges the LLM will follow literally. When you write “before X,” the LLM applies it everywhere it sees X — not just at the start of the workflow.

This is the second-order effect of prompt engineering. Every constraint needs to be read from the LLM’s perspective, not yours. What sounds unambiguous to you — “call this tool before any other” — may have a perfectly logical alternative interpretation that causes the agent to loop forever.

When writing constraints, specify three things:

1. How many times — “once”, “exactly once”, “at most once”
2. When — “as your first action”, not “before every…”
3. What to do next — “then move on” removes the question of what follows

The hallucination fix taught me that suggestions get ignored — you need MUST. The loop fix taught me that MUST isn’t enough — you need precision. One ambiguous word in a constraint and the agent follows the wrong interpretation faithfully, all the way to the iteration limit.

The agent did exactly what the prompt said. The prompt just didn’t say what I meant.