FoundationAgents/awesome-foundation-agents

What it claims

This is the reading-list companion to a specific survey, Advances and Challenges in Foundation Agents, not a standalone taxonomy of the maintainers' own devising. The repo is a flat, heavily-sectioned bibliography — hundreds of papers with [paper]/[code] links — organized to mirror that survey's structure. A Version 2 is flagged as "coming soon," so the current list is explicitly a moving target.

What the summary misses

The existing summary describes the constructive side (brain/perception/action, memory, planning, tool use, multi-agent, self-evolution) but omits that a large fraction of the list is agent safety and security — this is arguably its most distinctive slice for someone who already knows the agent-architecture literature. The visible portion carries dense sub-collections on:

• Jailbreaks (AutoDAN, GCG-style transferable attacks, embodied/POEX, multimodal)
• Prompt injection (indirect/tool-integrated, prompt-injection-to-judge, LLM-to-LLM "Prompt Infection")
• Data poisoning & backdoors (AgentPoison, sleeper agents, PoisonedRAG, RAG backdoors)
• Privacy (membership inference, training-data extraction, prompt-stealing/leakage, model-stealing)
• Physical/sensor attacks on embodied agents (DolphinAttack, LiDAR/GPS spoofing, adversarial ultrasound)

If your reason for opening this is "what can go wrong when I deploy an agent," that's here and the summary doesn't say so.

How to use it

It's a bibliography — clone or browse for citations; there is no framework or runnable artifact to adopt. Treat it as an index into a survey, useful mainly for a literature sweep on a named sub-topic (e.g. memory architectures: MemGPT, HippoRAG, A-MEM, Titans; or VLA models: RT-1/2, OpenVLA, π0, GR-2).

Limits

• Curation quality is uneven. The fetched text shows leftover formatting cruft — placeholder citation numbers (DINO-WM [358]), broken links ([[code][]], a truncated [paper](https://arxiv.org), see above back-references, and duplicated entries (Reflexion, MetaGPT, HippoRAG appear repeatedly). Some entries carry implausible venues (CogVLM: NeurIPS 2025, BERT 2018 with future-dated neighbors), so verify each citation before trusting it.
• V2 pending means links and coverage will shift.
• The page is paper links only — no summaries, no comparison, no synthesis per item; the actual argument lives in the linked survey, not here.
• Caveats: only the first ~60k of ~61k characters were retrieved. The truncation lands mid-way through the security section, so later sections (and whatever concluding material or contribution guidelines exist) are not visible and are not accounted for above.