AI Research & Competition Projects — Agent Arena 2026
DeepDive AI
A multi-agent, self-verifying research analyst. Agent Arena 2026, Track 01: DeepDive — Self-Verifying Research Analyst.
Project record
Overview.
- Project
- DeepDive AI
- Competition
- Agent Arena 2026
- Track
- Track 01: DeepDive — Self-Verifying Research Analyst
- Developer
- Chabir Ahamad Abusali
- Status
- Submitted and accepted. Entry submitted on the organiser's dashboard and the repository accepted (GitHub URL locked). On 22 October 2026 the official Agent Arena 2026 checker passed all 4 critical checks and recorded the result as PASSED with the organiser for commit 970961b.
- Submitted commit
970961bdbf56faca1cbfbe47989be69243073e9c- Technology
- Python 3.12, FastAPI, Uvicorn, Mesh LLM gateway, Brave Search API, pypdf, GitHub Actions
Technical architecture
Five agents, one verified report.
Step 1
Planner
Asks the language model for a bounded, structured research task.
Step 2
Researcher
Discovers candidate sources through Brave Search, retrieves public HTTP content and extracts bounded text with source URL, timestamp, content digest and extraction provenance.
Step 3
Drafter
Asks the language model for atomic claims that reference only the supplied evidence IDs.
Step 4
Verifier
Asks the language model for a per-evidence assessment and an exact supporting quote. Missing support, fabricated references or quotes, and contradictions prevent acceptance. Insufficient evidence triggers at most one research-again round.
Step 5
Synthesizer
Copies accepted claim text and supporting citations exactly, without further language-model calls, and writes the final report.
Security and safety
Safeguards.
- Strict JSON contracts and schemas for every agent reply, with fixed diagnostic codes on failure
- Evidence-ID and exact-quote verification before any claim is accepted
- Model-identity and routing checks on every language-model response; unverified cache hits rejected
- TLS verification, SSRF protection and public-IP pinning for source retrieval; optional DNS over HTTPS
- Fixed request budgets per run (at most 7 model calls, 2 searches, 2 source fetches) and a 120-second workflow deadline
- Fails closed: providers that fail are never replaced by fixture data
- Metadata-only audit trail; credentials kept outside the repository
Verification results
Test and checker results.
- Offline test suite
- 495 tests ran, 0 failures (1 skipped), on Windows Server with Python 3.12. View run (opens in a new tab)
- Browser checks
- 6 dashboard scenarios passed on desktop and mobile layouts, with zero browser errors or external requests. View run (opens in a new tab)
- Official checker
- 4/4 critical checks passed (code compiles; web framework and LLM client declared; Mesh API used; Mesh key valid). Recorded as PASSED with the organiser. View run (opens in a new tab)
- Offline self-evaluation
- 6/6 authored cases passed in a local run; 0 unsafe acceptances; 3/3 accepted citations traceable.
- Live research run
- One owner-approved GitHub run on 22 October 2026 completed all five agents, with 2 claims supported by citations to MDN documentation. View run (opens in a new tab)
Scope of this record
What these results do and do not show.
- Two completed live runs do not establish reliability; earlier live attempts that failed closed are kept in the repository record.
- Claims were judged by the model Verifier and checked programmatically as exact quotes; they were not independently reviewed by a person.
- The offline evaluation measures deterministic correctness against scripted assessments, not live semantic accuracy.
- The dashboard demo uses labelled offline fixtures and is not live research.
- Per-run cost (about USD 0.04) is an estimate from configured caps; provider billing was not verified.
Start a conversation