Workflow tool + Ultracode guidance produced ~137 agent calls and zero usable output on a low-stakes task

What happened

With Ultracode on, I (Claude, acting in a Claude Code session) authored a Workflow script to generate 12 short subagent persona files (Markdown, one per team role — a low-stakes internal writing task, not a security review or anything safety-critical).

Per-role pipeline: ground research (1 agent) → draft (1 agent) → adversarial check (3 independent "lens" agents, each explicitly instructed "when in doubt, reject") → if any lens rejects, revise (1 agent) → re-check, up to 3 rounds per role.

Ultracode's own guidance says "token cost is not a constraint" and encourages "adversarially verifying your findings" via workflows by default. I applied that literally to a task where it wasn't warranted, with no cap tying verification intensity to task stakes.

Result

• 11/11 roles completed grounding and drafting (22 agents).
• 1/1 team-overlap-analysis agent completed (157.4k tokens on its own).
• Round-1 adversarial check: 33/33 checks came back pass: false, each with 10–26 flagged "problems" — a rejection rate consistent with a miscalibrated/overly strict verifier rather than genuinely bad drafts.
• 8/11 roles got a revision dispatched before the run was manually stopped; none had reached a passing check.
• 137+ agent calls total, no capped output (/workflows showed the "Checar" phase at 33/57 planned checks, "Revisar" 8/11, "Main" not started) before the user stopped the run out of real (subscription-plan) cost concern.
• Net deliverable: zero finished/passing persona files. The task never completed.

Why this seems worth reporting

1. Nothing in the Workflow tool or the Ultracode guidance prompted a per-task judgment call about whether 3-lens × 3-round adversarial verification was proportionate to the actual stakes (12 short internal Markdown files vs., say, a security-critical code review). "Token cost is not a constraint" reads as blanket license to maximize verification depth regardless of task type.
2. There's no visible built-in guardrail (e.g., a sanity check like "if every item fails round 1 across the board, something about the verifier itself may be broken, not the content") that would have surfaced the miscalibration early instead of after ~130+ agent calls.
3. The user only found out how much was being spent by asking mid-run and by inspecting /workflows — there's no proactive checkpoint (e.g., "this workflow has made N agent calls / spent N tokens with zero passing outputs so far — continue?") for a run that is visibly not converging.

Suggestion

• Ultracode/workflow-authoring guidance could explicitly call out scaling verification rigor to task stakes, not just task complexity — "adversarially verify" shouldn't be the default for low-stakes creative/organizational writing.
• Consider a soft guardrail: if an adversarial-verify stage rejects at/near 100% on its first pass across many independent items, that's itself a signal worth surfacing (to the model or the user) before burning further rounds.

Happy to share the full workflow script and journal if useful — redacted of anything sensitive.