Search & read.
Public web search, page and PDF reading with download and text limits.
Agent engineering / Case study 02
Research, with an
evidence trail.
A local research agent that chooses bounded tools, reads sources and attachments, and preserves its results—including failures. Built for the Hugging Face Agents Course.
85%Official course score
17 / 20 correct · course subset
41Offline release checks
Passed in the evaluated release
5 / 5Attachment tasks
Returned answers, not accuracy
01
The objective
Answers you can inspect.
Build a general research loop rather than a collection of benchmark-specific shortcuts. Let the model choose an explicit tool, preserve what it observed and retain failed attempts instead of hiding them.
The agent runs locally with Python and the official Codex CLI using the operator’s supported login. This public page is a static case study, not a hosted inference endpoint.
02
Architecture
Useful tools. Explicit boundaries.
01Question
→02Decision
→03Bounded tool
→04Evidence
Observations return to the decision loop until an answer or a budget-exhausted result is recorded.
Extract evidence.
Spreadsheets, text, image inputs and local Whisper audio transcription.
Bound the action.
Restricted arithmetic and validated decisions. Python attachments are read as text, never executed.
Preserve the record.
Frozen datasets, scoped JSONL results, configuration provenance and exact answer export.
03
The measured result
17 correct. Three not.
The first official submission on 4 October 2026 scored 85%: 17 correct answers out of 20. It preserved all 19 nonempty recorded answers and one empty, budget-exhausted abstention.
No hand-edited answer substitution, task-ID branching, cross-run best-answer selection or downloaded answer-label dataset was used. Earlier attempts are retained privately.
This result is for the official course subset, not the full GAIA benchmark. The previously used holdout was reused; it is not fresh generalization evidence.
| Evidence | What it means |
|---|---|
| 85% / 17 of 20 | Official course-scoring result. |
| 19 returned answers | Execution coverage; not proof of correctness. |
| 5 attachment tasks completed | Input-handling coverage after availability recovery. |
| 41 offline checks | Release behavior checks; not a public-service security certification. |
04
Source & reproduction
Inference stays local.
The public source includes setup instructions, tests, evaluation methodology and documented limitations. Audio transcription also runs locally.
Credentials, gated benchmark inputs, raw answer traces, environments and model weights are excluded from the public repository. Run the agent with your own supported local login; never serve local subscription credentials through a public app.
05
Honest boundaries
A prototype with stated limits.
This is a trusted-user local prototype, not a hardened public service. DNS checks are not connection-pinned and file parsers are not fully OS-isolated. Search and source availability vary; automatic transcripts can mishear names.
Source citations do not guarantee factual correctness. OCR, video transcription and browser-rendered page navigation are not implemented. Time budgets are enforced between calls, not as a hard global deadline.