Citations & papers

"agentfootprint is not the inventor of these patterns. We are the substrate that lets you express them in 30 lines."

Patterns shipped as recipes

Every pattern in examples/patterns/ is a recipe over the 2-primitive + 4-composition substrate. The recipe is faithful to the paper's idea; some implementation details (the LLM, the prompts, the iteration count) are tunable. Each entry below cites the paper + links the recipe + notes the deviation.

ReAct (the default Agent loop)

Yao, Zhao, Yu, Du, Shafran, Narasimhan, Cao. "ReAct: Synergizing Reasoning and Acting in Language Models." ICLR 2023. arXiv:2210.03629

Recipe: the Agent primitive itself. No separate file — Agent.create({ provider, model }).tools([...]).build() IS ReAct. The framework owns the loop (think → act → observe → repeat).

Deviation: none — this is the canonical loop, with tool dispatch + result feedback identical to the paper.

Reflexion

Shinn, Cassano, Berman, Gopinath, Narasimhan, Yao. "Reflexion: Language Agents with Verbal Reinforcement Learning." NeurIPS 2023. arXiv:2303.11366

Recipe: examples/patterns/02-reflection.ts — Sequence(Agent, critique-LLM, Agent) wrapped in a Loop that exits when the critic emits DONE.

Deviation: the paper's full Reflexion includes "verbal RL" (memory of past mistakes informing future attempts). Our recipe is the propose-critique-revise loop only; long-horizon verbal-memory is left to the consumer (or a defineMemory({ type: NARRATIVE }) layer).

Tree-of-Thoughts (ToT)

Yao, Yu, Zhao, Shafran, Griffiths, Cao, Narasimhan. "Tree of Thoughts: Deliberate Problem Solving with Large Language Models." NeurIPS 2023. arXiv:2305.10601

Recipe: examples/patterns/05-tot.ts — Parallel(Agent × N) + LLM-rank merge.

Deviation: paper's full ToT includes BFS / DFS search over the thought tree with backtracking. Our recipe is one-level fan-out + LLM rank — covers the "Self-Consistency-with-explicit-reasoning" use case. Multi-level search is composable via nested Parallel + Loop, not currently shipped as a recipe.

Self-Consistency

Wang, Wei, Schuurmans, Le, Chi, Narang, Chowdhery, Zhou. "Self-Consistency Improves Chain of Thought Reasoning in Language Models." ICLR 2023. arXiv:2203.11171

Recipe: examples/patterns/01-self-consistency.ts — Parallel(Agent × N) + majority-vote merge.

Deviation: none meaningful — same idea, expressed as a Parallel + a deterministic merge.

Debate

Du, Li, Torralba, Tenenbaum, Mordatch. "Improving Factuality and Reasoning in Language Models through Multiagent Debate." ICML 2024. arXiv:2305.14325

Recipe: examples/patterns/03-debate.ts — Loop(Agent × 2 + judge).

Deviation: paper varies N agents, M rounds, judge configurations. Our recipe is two debaters + one judge + bounded rounds — the simplest form. N-agent variants compose via wider Parallel.

Map-Reduce

Dean, Ghemawat. "MapReduce: Simplified Data Processing on Large Clusters." OSDI 2004. Original paper

Recipe: examples/patterns/04-map-reduce.ts — Parallel(Agent × N) + LLM-merge.

Deviation: not really a deviation — Map-Reduce in the agent context is fan-out-then-merge, which is what Parallel does. The original paper is about distributed computing; the LLM-agent flavor is a faithful translation.

Swarm

OpenAI engineering. "Swarm: An ergonomic, lightweight multi-agent orchestration framework." 2024. GitHub

Recipe: examples/patterns/06-swarm.ts — swarm({ agents, route, maxHandoffs }) over Loop.

Deviation: OpenAI's Swarm uses LLM-driven routing (each agent decides the handoff). Our recipe is sync-deterministic-route by default (the route function is a pure JS function). LLM-driven routing is composable via Conditional with an LLM predicate, but the deterministic flavor is what production teams actually want for predictable cost + observability.

Skills (Anthropic Agent SDK)

Anthropic engineering. "Agent Skills." 2024. Documentation

Recipe: defineSkill({ id, description, body, tools }) with auto-attached read_skill tool. See Skills, explained for the full essay on how the implementation differs from Anthropic's Claude-first design.

Deviation: agentfootprint's Skills work cross-provider via tool-result delivery (recency-first, protocol-level guarantee). Anthropic's SDK uses system-prompt anchoring (Claude-trained-correct, drifts on other providers). Our v2.4 Phase 4 ships surfaceMode: 'auto' to pick per provider.

Causal memory (decision-evidence persistence)

No prior paper. This is agentfootprint-original, building on:

• footprintjs's decide() / select() evidence-capture primitives — see footprintjs docs
• Hippocampus-inspired causal memory frameworks — broadly inspired by cognitive-architecture work on episodic vs causal memory; specific paper inspirations include LIDA (Franklin et al.) and ACT-R (Anderson) but the implementation is structural, not theory-driven

The closest published analog is MemGPT-style snapshot memory (Packer et al., arXiv:2310.08560), but MemGPT snapshots message history; we snapshot decision evidence. Different shape, different downstream economics (audit + cheap-model triage + training data).

See Causal memory deep-dive for the snapshot shape + worked replay.

Foundational frameworks we build on

• footprintjs — the flowchart-pattern execution engine. The "framework owns the loop" property comes from footprintjs's DFS traversal. GitHub.
• Hexagonal Architecture (Alistair Cockburn, 2005) — provider adapters as ports. Outer-ring isolation.
• GoF Patterns — Adapter pattern for every provider/store integration. Decorator pattern for resilience composition.

Augmented LM framing

The unifying mental model — "Skills, RAG, Memory, Instructions, Tools are all augmentations of the LLM call" — maps to:

Mialon, Dessì, Lomeli, Nalmpantis, Pasunuru, Raileanu, Rozière, Schick, Dwivedi-Yu, Celikyilmaz, Grave, LeCun, Scialom. "Augmented Language Models: a Survey." TMLR 2023. arXiv:2302.07842

This survey's framing — every external input as an augmentation — is the conceptual root of our Injection primitive. We just gave it a concrete shape ({ trigger, inject, flavor } — when to activate, what to contribute across slots, and an observability tag) and a typed event channel.

How to cite agentfootprint

If you build on agentfootprint in academic work, please cite:

@software{agentfootprint,
  title = {agentfootprint: Context engineering, abstracted},
  author = {Sanjay Krishna Anbalagan},
  year = {2026},
  url = {https://github.com/footprintjs/agentfootprint},
}

Next steps

• Manifesto — the framework's perspective on what it is + isn't
• Skills, explained — the strongest essay in the docs, with full cross-provider correctness analysis
• Causal deep-dive — the differentiator, made tangible