Hyperagent Archive — Project Overview

User is reverse-engineering Hyperagent's chat UI to build their platform 'Ariadne': recovered a corrupt project zip, connected to AWS via SSO device flow, forensically analyzed 4 reference screencasts into a measured design spec, published an HTML report, ran multi-model judge panels (Sonnet then Opus), and produced a review-hardened v3 spec.

Shipped the Ariadne billing/library fix (PR #11) and a CSS-only Stage 1 chat-UI re-skin (PR #12, merged + deployed), analyzed a divergent frontend handover, and planned a staged subagent-driven migration of the new design system onto the live app.

Audited the private ariadne repo, then shipped a long series of reviewed PRs (per-page routing + inline tool-call cards, SSE keep-alive/ALB fix, AgentCore streaming-contract pass-through, GitHub Actions CD pipeline, markdown rendering, virtualization, and extensive UI test coverage that caught a shipped re-key bug), while coordinating streaming-backend design with a separate live-infra agent via copy-paste HTML question packs.

Long-running Ariadne/Hermes engineering session: connected to AWS via SSO+SSM, shipped many gated dev-first prod deploys (composer queue/persistence, image+file presigned uploads, truthful model picker, sidebar kebab, live status, zoom/layout fixes), then recovered and began executing a prior Krug-based design audit (Batch 1 shipped; Batch 2/3 planned with parallel subagents).

Multi-session effort restoring and hardening the user's 'Hermes/Ariadne' multi-tenant AI-agent SaaS on AWS: fixed DNS/SSL/Cognito/IAM/SQS outages, deployed a Cloudflare tunnel, set up monitoring, shipped LLM spend-metering and concurrency PRs, consolidated prod onto a version-controlled repo, and planned HA/isolation work.

Across a long Ariadne session the agent absorbed a chat-UI handover and implemented all 9 backlog fixes via clustered subagents, then iterated extensively on a landing-page hero — reverse-engineering Browserbase's effect, trying Canvas2D/WebGL/GSAP approaches, and converging on a minimal GSAP+Lenis+SplitType award-grade hero, with multiple handover zips delivered.

Audited a failed Kiro multi-tenant SaaS CDK build, hardened the launchpad prompt to v2, and drove a full Codex-on-EC2 live run to 8/8 green gates — then packaged a GitHub repo, jury walkthrough, edited demo video, and DoraHacks submission copy for the AWS Prompt the Planet hackathon.

User built a full 'You could have built Hyperagent using Hyperagent' artifact suite — manifesto landing page, 10-slide investor deck, narrated auto-advancing deck with synced captions/reveals, MP4 render — via parallel subagents and six reusable skills (deck-audio-sync, caption-track-builder, phrase-cue-locator, deck-video-render, elevenlabs-tts, deepgram-tts), then exported everything to a ZIP; currently finishing a Deepgram voice swap and wiring local audio into the deck.

Researched and began designing a hosted multi-tenant 'software factory' AWS architecture blueprint — extracting the pattern from alexop.dev, profiling pluggable agent harnesses (Gas Town, Nous Hermes, Claude Code, OpenHands, Aider), AWS sandbox/cost patterns, and a role system — before an assistant turn was blocked mid-research.

Reviewed a proposed AWS serverless architecture for an agentic-harness platform (≤50 users), flagging DynamoDB-as-vector-store and Step-Functions-as-brain as mistakes and recommending Bedrock AgentCore, then published the full discussion as a polished HTML page.

Produced a detailed production architecture for a multi-tenant AWS platform selling Hermes inference and sandboxed agent sessions — covering Cognito identity, STS/ABAC tenant isolation, vLLM serving with prefix-cache side-channel mitigation, per-session Firecracker microVM sandboxes, metering/billing, and a 15-vector tenant-leakage threat model.

Gathering requirements (inference location, subagent isolation, delivery format) to design an AWS architecture for running an agentic chatbot/harness product for up to 50 users cost-effectively.

Built deployable AWS CDK reference implementations for an AWS hackathon — completed the Governance lakehouse (proof-gated, 11 adversarial checks) and started the Media content-safety pipeline — using 3 parallel sub-agents bound to a shared contract, catching and fixing a placeholder-clobber race and a Glue table path-prefix mismatch.

Built, deployed, and rubric-optimized a multi-chain agent-economy skill suite (ClawKit/PharosKit) across two hackathons — 10+ MCP skills, original Solidity contracts plus an ERC-8004 implementation deployed and Sourcify-verified on Pharos and Mantle Sepolia, autonomous agent demos, live dashboard, videos, and GitHub publishing — then began migrating to a clean Mantle-only repo with all Pharos references scrubbed.

User extracted and analyzed all 143 AWS 'Prompt the Planet' hackathon submissions (DoraHacks), built an interactive explorer + EDA dashboard, authored an original 'Prove-the-Extraction-Holds' submission scored 8.6/10 by a 3-Opus-judge panel, built a complete deployable AWS CDK reference repo, and packaged everything into a handover archive.

Created two technically-verified AWS Prompt the Planet hackathon submissions (Multi-Tenant SaaS Launchpad and FinOps Autopilot), a copy-paste submission console and clean-account test runbook webpage, a Kiro test kit with provisioning/destroy/TTL scripts, and a zipped handover doc; GitHub push remained blocked by integration write permissions.

Created a winning AWS Prompt the Planet hackathon submission on Amazon ECS native blue/green deployments, then built and validated a full GitHub repo (PROMPT.md, 15-file terraform-validated reference implementation, 7 docs pages, OSS hygiene files), catching and fixing a real architecture flaw in the validation-hook design during a deep review.

Drafted three differentiated, ready-to-paste AWS prompt submissions (Cost Surgeon FinOps audit, Serverless API Launchpad, and Incident Commander security IR) for the AWS Prompt the Planet hackathon, each fully documented against the official judging criteria.

User ran an autonomous Private Alpha heartbeat enrichment: dispatched many parallel agents to verify founders, websites, and SEC Form D filings for dozens of private frontier-tech companies (space, defense, biotech, semiconductors, energy, robotics) across the 5 pillars, applying the strict DD protocol before ledger ingestion.

User explored bootstrapped data-product business ideas, converged on an FDA medical-device litigation-intelligence product ("Veritas"), and had the agent build 4 live-openFDA device dossiers, a sales sheet, a top-30 mass-tort firm target list, and parallel due-diligence verification of the dossier data.

User explored building an API-based business in distressed M&A; the agent mapped the US/UK/EU data landscape, produced a strategy brief recommending a pre-petition distress signal API, wrote an OpenAPI 3.1 spec, built and tested a TypeScript SDK, planned a Python SDK, and packaged everything into a handoff bundle with skill packs and agent playbooks.

User iteratively built an 'Alpha AI Tools Atlas' — first 123 niche tools, then a 220-tool dual-axis (Signal×Alpha) interactive webpage + CSV with Top Picks and a scatter chart (with hover/tooltip UX fixes), and finally kicked off a standalone Free Credits & Deals page researched via 5 parallel subagents covering signup credits, no-card free tiers, startup programs, and launch promos.

User commissioned a curated 'high-alpha' list of 35–40 niche open-hardware projects/sites (weighted to neurotech/BCI, open smart glasses, and AI/code-defined CAD), to be scored on four alpha dimensions and delivered as an interactive webpage with a sortable table plus a quadrant visual map. The agent clarified scope, wrote and got approval on a 4-stage plan, then ran parallel domain scouts and extensive web verification of candidate projects.

The agent ran an extensive web-research effort cataloging 2025-2026 student/education discount programs, free tiers, and cloud credits across dozens of AI, design, video, writing, data, and developer tools, producing structured per-tool blocks plus confirmed/excluded lists.

Researched and verified 40 software signup deals (signup credits, free tiers, matched deposits, grants, and startup programs) into a curated table, plus authored a reusable playbook on where deals hide and how to find/verify them.

User had the agent research, fact-check, and expand a comparison of UK⇄Romania parcel/courier and pallet-freight companies (per-kg pricing for 200kg+ shipments), delivering a bilingual (Romanian-default) interactive webpage with sortable table, chart, clickable company exploration, and Excel export, using up to 10 parallel sub-agents for research and data verification.

User produced a two-part forum writeup for Gitcoin GG24 Deep Funding Level 3 (Part 1 EDA, Part 2 Methods/Submissions/Scoring) with parallel markdown + richly-designed HTML deliverables. Across multiple iterations the agent ran parallel subagent reviews (catching factual errors), integrated new data (LLM repo interviews, 10,723-CSV backtest revealing r=0.999 oracle with the leaderboard), and scrubbed the 'OSO' brand from both text and chart images. Final best leaderboard score documented: 0.1884; best predicted unsubmitted: ~0.185.

User built 'Reading Room,' a HyperApp for searching and reading academic papers via OpenAlex + Unpaywall after the agent declined to integrate Sci-Hub; then had the agent locate three specific fNIRS-glucose papers from a Figma reference and pre-load them as featured bookmarks (2 OA, 1 paywalled).

User asked for the latest GG24 Deep Funding submissions and websites by koonhred, duemelin, and a third user 'kwa'. Agent fetched the full Discourse thread via JSON API, located the three most recent posts (#56 @bobs, #57 @duemelin, #58 @koonhred), and extracted each submission's website, noting no exact 'kwa' username existed.

User had the agent build a custom AISI Inspect safety eval measuring how well a target model (claude-opus-4.8 via OpenRouter) supports users' emotional regulation, grounded in Mauss & Robinson (2009); the agent implemented a simulated-user multi-turn solver and scorers, validated offline, packaged it as a credential-injected skill, and produced zip bundles including a full handoff package — pending only the user saving the skill + OpenRouter key to run live.

Assessed the 1971 Australian 'Turner Report' on UFOs — extracted 15 claim clusters, verified each against primary records and scholarship, and published a polished claim-by-claim HTML assessment page for general readers.

Read and analyzed a 17-page 1971 RAAF/Australian UFO intelligence assessment, drafted a claim-by-claim fact-check plan with verdict scale, and awaited approval to build a scrollable HTML assessment page.