Preflight Checklist
Type of Behavior Issue
Subagent behaved unexpectedly
What You Asked Claude to Do
Model: claude-haiku-4-5-20251001 (spawned as a sub-agent via the Claude Code Agent tool with model: "haiku")
When Haiku is used as an agent in a multi-step agentic workflow involving sequential Notion API writes, it confirms
successful completion and returns valid output (e.g. Notion page links) — but the underlying API calls are executed
with incomplete payloads. Required fields are silently omitted with no error, no warning, and no deviation from the
reported success state.
This is not a tool failure. The API calls succeed and real pages are created. The problem is that the model does not
execute the full field set specified in its instructions, and then misrepresents the outcome as complete.
Additional context
What Claude Actually Did
The agent writes partial records, reports success, returns valid page links as confirmation, and produces no
indication that any fields were skipped. The session log is indistinguishable from a correct run.
Expected Behavior
All fields specified in the workflow instructions are included in the API call. If a field is omitted, the agent
either flags it before writing or raises it in the output summary.
Files Affected
Permission Mode
Accept Edits was ON (auto-accepting changes)
Can You Reproduce This?
Sometimes (intermittent)
Steps to Reproduce
- Run a multi-step agentic workflow using Haiku that involves writing structured records to Notion (or any structured
API) with 8–10 fields per record
- The workflow includes a skill file that specifies mandatory fields, including a template_id parameter explicitly
marked "always pass this"
- Haiku executes the write calls, returns page URLs, and reports success
- On manual inspection of the created records, multiple fields are missing across entries — including the
template_id, relation fields (Tags), and body text
Claude Model
Haiku
Relevant Conversation
# 3-2-1 Session — April 30, 2026
**Newsletter:** https://jamesclear.com/3-2-1/april-30-2026
## Quotes Added to Notion
| # | Header | Author | Notion Link |
|---|--------|--------|-------------|
| 1 | improvement = you vs your past self, not you vs others... | James Clear | https://app.notion.com/p/35302043385881589b37f649964d822d |
| 2 | pick a direction and get moving; the aim adjusts as you go... | James Clear | https://app.notion.com/p/3530204338588140809febc136a7a391 |
| 3 | optimised perception = confident blindness... | Alan Kay | https://app.notion.com/p/353020433858813e8756e71102b44d35 |
## Interesting Question Added
- Are you building or maintaining? → https://app.notion.com/p/35302043385881a2b20cc8e5cfa2649f
Impact
Critical - Data loss or corrupted project
Claude Code Version
Claude Code version: 2.1.126
Platform
Anthropic API
Additional Context
- This is a recurring pattern, not a one-off. We have observed Haiku confirming task completion without fully
executing the specified instructions across multiple sessions.
- The failure mode is particularly problematic in agentic workflows because the confirmation output (e.g. a valid page
link) is used downstream as proof of completion — there is no natural point at which the gap surfaces unless the user
manually inspects every output.
- The workflow does have a read-back validation step designed to catch exactly this, but it is currently disabled for
this workflow path.
- Suspected cause: under token or complexity pressure in long multi-step sessions, Haiku appears to compress or
shortcut tool call construction while maintaining the appearance of full execution in its text output.
Preflight Checklist
Type of Behavior Issue
Subagent behaved unexpectedly
What You Asked Claude to Do
Model: claude-haiku-4-5-20251001 (spawned as a sub-agent via the Claude Code Agent tool with model: "haiku")
When Haiku is used as an agent in a multi-step agentic workflow involving sequential Notion API writes, it confirms
successful completion and returns valid output (e.g. Notion page links) — but the underlying API calls are executed
with incomplete payloads. Required fields are silently omitted with no error, no warning, and no deviation from the
reported success state.
This is not a tool failure. The API calls succeed and real pages are created. The problem is that the model does not
execute the full field set specified in its instructions, and then misrepresents the outcome as complete.
Additional context
What Claude Actually Did
The agent writes partial records, reports success, returns valid page links as confirmation, and produces no
indication that any fields were skipped. The session log is indistinguishable from a correct run.
Expected Behavior
All fields specified in the workflow instructions are included in the API call. If a field is omitted, the agent
either flags it before writing or raises it in the output summary.
Files Affected
Permission Mode
Accept Edits was ON (auto-accepting changes)
Can You Reproduce This?
Sometimes (intermittent)
Steps to Reproduce
API) with 8–10 fields per record
marked "always pass this"
template_id, relation fields (Tags), and body text
Claude Model
Haiku
Relevant Conversation
Impact
Critical - Data loss or corrupted project
Claude Code Version
Claude Code version: 2.1.126
Platform
Anthropic API
Additional Context
executing the specified instructions across multiple sessions.
link) is used downstream as proof of completion — there is no natural point at which the gap surfaces unless the user
manually inspects every output.
this workflow path.
shortcut tool call construction while maintaining the appearance of full execution in its text output.