| You did | Do instead | Why |
|---|---|---|
| Ran an agent to look up one fact | Search with include_answer | Seconds and cents instead of minutes and dollars. See Choosing the Right Tool |
| Gave the agent one known URL | Extract, or the site’s Extract Template | A known page is read deterministically. An agent searches for what you already have |
| Created a new agent for every question | One agent by agent_name, many runs | The tuned skill, goals and sources are the asset. See Running at Scale |
| Asked one run to crawl a whole catalog | 50 to 100 rows per run, scoped by section, then merge | One x-high run over a full manufacturer tree returned 78 items; scoped runs returned 200 or more each |
| Asked for “all”, “every”, “comprehensive” on the open web | A row cap and a stop rule; split by region or category | The agent stops when the budget ends, not when the task is done. See Writing the Task |
| Put twelve entities in one research prompt | One entity per run | Twelve products returned five rows covering two, and the run said completed |
| Wrote goals as caveats and the skill as a feature list | Skill as a playbook with steps; goals as checkable requirements | Rewriting one skill into a procedure lifted filled cells from 68% to 99% |
| A 25-field schema with no descriptions | 5 to 8 described fields; a second agent for the rest | Wide undescribed schemas are the top cause of partial output. See Output Schema |
Nested arrays in an enrichment schema, or $ref | Flat object per row; inline definitions | Nested arrays exhaust the write budget; $ref fails at run start |
| Required fields the web cannot fill | required on identity fields only; everything else nullable | The agent invents a value or drops the row |
Asked for confidence or reasoning fields in the schema | Leave them out; read the trust report | The report already grades every path. Self-grading is a worse copy |
| Built on LinkedIn, Instagram, Zillow or a login-walled site | Drop the field, use a template, or ask about social data access | Login and anti-bot walls return thin or padded results. See Controlling Sources |
| Hunted personal emails and mobile numbers | Ask for the URL of the published contact page; forbid pattern-built emails | Rarely public; invites construction |
medium effort with a structured schema | Free text at medium, or high with the schema | A third of those runs ship partial output. See Effort and Cost |
| Raised effort to fix a vague task | Fix the task text and schema first, effort last | Higher tiers buy sources and time, not clarity |
| Pasted 300 prior results into the prompt to dedupe | The prior rows in input_data with “exclude these” | Keeps the prompt short and the exclusion exact |
| Copied a gallery template and ran it unchanged | Read the copied agent’s goals and sources first | Copies have come across with empty goals and sources |
Trusted completed, or the confidence grade, as proof | Count rows against asked, check fill, open five citations | Truncated and padded runs both complete with a high grade. See Evaluating Results |
| Skipped canaries on a big batch | Ten known inputs first, including two that should return nothing | Catches wrong-entity and fabrication before 1,000 runs repeat it |
| Called synchronously with a 5-minute timeout | Poll or stream events; allow 30 minutes at high and x-high | Nine in ten runs finish sooner; the tenth does not |
Next Steps
Choosing the Right Tool
Before the first run
Writing the Task
Input, skill, goals
Evaluating Results
After the run