Resources

Annotated resource library

The best writing on agents and loops, selected for practical value and kept current by this site's own discovery and link-check loops.

How entries are chosen →

Filter the library

Filtering is implemented as a small progressive-enhancement script: with JavaScript off, the complete build-time rendered list remains readable and grouped by topic.

Showing 40 resources.

current40current
needs review0needs review
link dead0link dead
replaced0replaced

benchmarks

currentpaperadvanced

SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering

Yang et al. · added 2026-07-07

Live link, checked 2026-07-07

Introduces SWE-agent and the agent-computer interface (ACI) — the idea that interfaces designed for LMs outperform human-designed shells. State-of-the-art on SWE-bench at publication and the conceptual basis for how most coding-agent harnesses are now built. Read for the ACI argument; the benchmark numbers age, the design lesson does not.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Peer-reviewed (NeurIPS 2024); the ACI concept is a durable harness-design lesson and the basis for coding-agent interfaces. The benchmark framing also grounds the benchmarks topic. Distinct from the tools/harness essays because it is the research foundation.
  • benchmarks
  • coding-agents
  • harnesses
currentrepoadvanced

SWE-bench: Can language models resolve real-world GitHub issues?

Jimenez et al. (Princeton NLP) · added 2026-07-07

Live link, checked 2026-07-07

across 12 Python repos, graded by human-written unit tests. The reference benchmark for coding agents and the reason SWE-agent and its successors exist. Use it to understand what "verified" means for a coding-agent loop.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Princeton NLP, the benchmark the field uses. Execution-based verification is exactly the maker/checker model Loopmaster teaches. Distinct from SWE-agent as the benchmark vs the harness.
  • benchmarks
  • coding-agents
  • verification
currentpaperadvanced

τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains

Yao et al. (Sierra Research) · added 2026-07-07

Live link, checked 2026-07-07

Benchmarks agents in dynamic conversation with a simulated user under domain-specific API tools and policy guidelines — i.e. can the agent follow the rules, not just call the tools. Reveals inconsistencies in function-calling adherence that static benchmarks miss. The right benchmark to cite when verification means "did it obey policy under pressure," which is Loopmaster's guardrails pillar.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Sierra Research (Yao of ReAct). The tool-agent-user + policy framing is a distinct benchmark angle that tests guardrails, not just capability. Fills the guardrails-verification gap in the library.
  • benchmarks
  • guardrails
  • patterns

coding-agents

currentdocintermediate

OpenAI Codex

OpenAI · added 2026-07-07

Live link, checked 2026-07-07

OpenAI's coding-agent product and CLI. Relevant to Loopmaster as a reference coding-agent harness and as a backend the templates can orchestrate. More product surface than teaching resource; read it to understand the shape of a shipped coding agent.

Why it's here
Provenance
seed
Rubric score
7/10
Rubric note
Authoritative as a vendor product page, but it is product/marketing-leaning with less transferable technique than the Agents SDK docs. Lower durability and actionability for the same reason. Included as a reference coding-agent backend, not a teacher.
  • coding-agents
  • frameworks
currentrepointermediate

Aider

Paul Gauthier · added 2026-07-07

Live link, checked 2026-07-07

Open-source terminal coding agent that pairs with any git repo. A clean, hackable formats, repo-map context, and the commit/PR loop. Worth reading the source even if you use a different coding agent in production.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Long-running, well-maintained OSS coding agent with a distinctive repo-map/edit-format approach. Durable because the harness design lessons outlive model swaps. Uniquely readable as a reference harness; directly actionable.
  • coding-agents
  • frameworks

context-engineering

currentessayadvanced

Effective context engineering for AI agents

Anthropic · added 2026-07-07

Live link, checked 2026-07-07

Treats context as a finite resource and lays out strategies for curating and managing grounded as a run grows. The companion piece to harness engineering — if the harness is the shape, context is the fuel. Advanced but worth it for long-running loops.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Anthropic engineering on a topic (context management) that is the other half of harness engineering and not covered by the other entries. Concrete strategies; directly relevant to long-running and multi-context-window loops.
  • context-engineering
  • harnesses

eval-tools

currenttooladvanced

Inspect AI

UK AI Safety Institute · added 2026-07-07

Live link, checked 2026-07-07

OSS framework (MIT) from the UK AISI that has become the de facto standard for agent safety evals — Apollo, METR, Anthropic, and OpenAI publish risk evaluations on it. Built around task/solver/scorer decorators; one run produces a serialized log with both scores and trajectories. The right starting point if you need reproducible, sandboxed agent evaluation rather than a fast CI regression suite.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Government lab tool that is the consolidation point for agentic safety eval. Reproducible logs and trajectory capture are exactly the verification machinery Loopmaster advocates. Distinct from the CI-regression tools (Promptfoo, DeepEval).
  • eval-tools
  • verification
  • benchmarks
currenttoolintermediate

Promptfoo

Promptfoo · added 2026-07-07

Live link, checked 2026-07-07

`promptfoo eval`, get a comparison table plus red-team/adversarial coverage. Best fit for fast CI regression suites, model A/B, and security probing across providers. Pair with Inspect AI for safety campaigns and DeepEval for Python-native metric gates.

Why it's here
Provenance
seed
Rubric score
9/10
Rubric note
Widely adopted OSS with a distinct CLI/YAML ergonomics and best-in-class red-teaming. Durable as the CI-regression default. Overlaps the eval-tools cluster on the general eval angle but its declarative-CI shape is materially different from Inspect/DeepEval.
  • eval-tools
  • verification
  • guardrails
currentrepointermediate

DeepEval

Confident AI · added 2026-07-07

Live link, checked 2026-07-07

metrics (G-Eval, faithfulness, relevancy, hallucination) with thresholds; fail builds in CI. Best fit for Python teams who want evals as code in their existing test suite. Ships RAG and agent-task metrics; pair with Promptfoo for declarative comparison.

Why it's here
Provenance
seed
Rubric score
9/10
Rubric note
Maintained OSS, the pytest-shaped eval tool. Durable because the metric set and the pytest integration are stable abstractions. Uniqueness is moderate vs Promptfoo (the two are the standard pair) but the Python-native shape is distinct.
  • eval-tools
  • verification
currentessayintermediate

Your AI Product Needs Evals

Hamel Husain · added 2026-07-07

Live link, checked 2026-07-07

Argues the root cause of unsuccessful AI products is a failure to build robust eval domain expert as the single quality decision-maker, then build automated evaluators for repeated failure modes. The mindset piece that makes the eval tools worth using.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Husain is the most cited practitioner voice on LLM evals. The error-analysis-first methodology is distinct from the tool-focused entries and is the meta-lesson the tools serve. Directly actionable.
  • eval-tools
  • verification
currentessayintermediate

'It's Hard to Eval' Is a Product Smell

Hamel Husain · added 2026-07-07

Live link, checked 2026-07-07

If an AI artifact is hard for you to verify, it is probably hard for users too — and that is a product-design problem, not an evals problem. Walks through designing products so the output is checkable (provenance, progressive disclosure, smaller accept/reject units) before you reach for automated graders. The missing link between "build the loop" and "verify the loop."

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Same authority as the evals post. The "design for verifiability" angle is distinct and directly relevant to loop design — the output shape determines what verification is even possible. Not duplicated by the tool entries.
  • eval-tools
  • verification
currentessayadvanced

How do I evaluate agentic workflows?

Hamel Husain · added 2026-07-07

Live link, checked 2026-07-07

box, then step-level diagnostics (tool choice, parameter extraction, error handling, context retention, efficiency) using transition-failure matrices. The transition-matrix technique turns overwhelming trace review into a focused debugging map. The most directly applicable eval write-up for agent loops in the library.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Co-authored with Shreya Shankar from teaching 700+ engineers. The transition-matrix method is a concrete, agent-specific eval technique not covered by the tool entries or the general evals post. The most on-point eval resource for Loopmaster's verification pillar.
  • eval-tools
  • verification
  • patterns
currentessayadvanced

A review of OpenAI's o1 and how we evaluate coding agents

Cognition · added 2026-07-07

Live link, checked 2026-07-07

feedback, evaluator agents that judge outcomes with shell/edit/browse tools, and explicit modeling of user intent to auto-detect deviations. The "evaluating agents with agents" and "critiquing is easier than solving" observations are durable lessons for anyone building a verification loop for coding agents.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Cognition (Devin team), production coding-agent evals. The evaluator-agents and intent-deviation-detection patterns are distinct from the academic benchmarks and the general eval essays. Directly relevant to coding-agent verification.
  • eval-tools
  • benchmarks
  • coding-agents
currentrepointermediate

OpenAI Evals

OpenAI · added 2026-07-07

Live link, checked 2026-07-07

models. Now overtaken by Promptfoo and Inspect AI for greenfield projects, but the public registry of 100+ eval definitions is a useful reference for how others structure evaluations. Treat as a reference library, not the default starting tool.

Why it's here
Provenance
seed
Rubric score
7/10
Rubric note
OpenAI-authored, authoritative. Lower durability and actionability because newer tools overtook it; the value is the reference eval library. Included as the reference point, not a recommended default — the annotation says so honestly.
  • eval-tools
  • benchmarks

frameworks

currentrepointermediate

LangGraph

LangChain · added 2026-07-07

Live link, checked 2026-07-07

Graph/state-machine approach for long-running durable agents — the production-grade sibling of LangChain's agent abstractions. Relevant to Loopmaster as a candidate backend for templates that need durable state, human-in-the-loop interrupts, and branching control flow. Reference, not a teaching resource.

Why it's here
Provenance
seed
Rubric score
9/10
Rubric note
Maintained by LangChain, widely adopted in production. Durable as a framework reference though the API evolves. Actionable as an installable backend; uniqueness is moderate because it overlaps the broader LangChain 1.0 entry.
  • frameworks
  • orchestration
currentrepointermediate

CrewAI

CrewAI · added 2026-07-07

Live link, checked 2026-07-07

then compose them into a crew. A reasonable default for orchestrator-workers patterns where role specialization matters more than fine-grained state. Loopmaster treats it as an execution backend, not a competitor.

Why it's here
Provenance
seed
Rubric score
9/10
Rubric note
Popular production framework with a distinct role-based abstraction. Durable as a backend option; overlaps with other framework entries on the general agent-loop angle but its role model is a meaningfully different shape.
  • frameworks
  • orchestration
currentrepointermediate

Microsoft AutoGen

Microsoft · added 2026-07-07

Live link, checked 2026-07-07

Multi-agent conversation/workflow framework from Microsoft Research with a strong research lineage. Good for experimentation with agent-to-agent conversation patterns and group chat topologies. Heavier than the others; pick it when the multi-agent conversation itself is the thing you are studying.

Why it's here
Provenance
seed
Rubric score
9/10
Rubric note
Microsoft Research, credible. Durable abstraction (multi-agent conversation) that is a distinct shape from role-based or graph-based frameworks. Actionable as an installable library; modest uniqueness overlap with the framework cluster.
  • frameworks
  • orchestration
currentdocintermediate

OpenAI Agents SDK (Python)

OpenAI · added 2026-07-07

Live link, checked 2026-07-07

the tool-calling loop. Useful both as a backend and as a reference for how a major vendor factored the agent loop into primitives. The guardrails-as-first-class-object design is worth reading even if you never deploy it.

Why it's here
Provenance
seed
Rubric score
9/10
Rubric note
OpenAI-authored, authoritative. Durability is lower because the SDK tracks API changes and the abstraction is young. Uniqueness is high for the guardrails-as-object framing which other frameworks treat more loosely.
  • frameworks
  • loop-engineering

guardrails

currentessayadvanced

Giving Antigravity Agents Safe Write Access — Production Permission Boundary Design

Antigravity Lab · added 2026-07-07

Live link, checked 2026-07-07

operation model (read / sandbox-write / prod-write / destructive), a single executeWithBoundary gate for every write, dry-run + diff + approval + reverse-SQL per action, and per-agent tier caps in a registry. The most concrete permission-model write-up in the library; directly applicable to Loopmaster's guardrails pillar and loop.yaml permissions.

Why it's here
Provenance
seed
Rubric score
9/10
Rubric note
A production write-up from a less-widely-known author (authority 1) but extremely are reusable design. Uniquely fills the production-permission gap; durable because the patterns are vendor- and model-agnostic.
  • guardrails
  • maintenance
currentessayintermediate

AI Agent Tool Permissions (With Code)

Agent Patterns · added 2026-07-07

Live link, checked 2026-07-07

scope credentials to tenant + environment, treat approval as a first-class state, and a practical capability-token shape. The "prompt injection is a permissions problem, not a prompt problem" framing is the right mental model. Pair with the permission-boundary post for the production version of these ideas.

Why it's here
Provenance
seed
Rubric score
8/10
Rubric note
Practitioner-patterns site; authority is moderate (1) as a multi-author resource. Durable because the least-privilege patterns are stable. Overlaps the permission-boundary post on the deny-by-default idea but is the lighter, more introductory version.
  • guardrails

harnesses

currentessayintermediate

Agent Harness Engineering

Addy Osmani · added 2026-07-07

Live link, checked 2026-07-07

Argues the harness — model plus tools, prompts, hooks, subagents, memory, and guardrails — is the actual agent product, and that the raw model is a component. Complements the Read both; together they are the Loopmaster thesis in two posts.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Same authority as the loop-engineering post. The harness framing is distinct from the loop framing and together they cover the two things Loopmaster teaches. Concrete about what makes a harness good vs a raw model.
  • harnesses
  • loop-engineering
currentessayintro

How coding agents work

Simon Willison · added 2026-07-07

Live link, checked 2026-07-07

(especially code execution), and the model+tools-in-a-loop core. The right first read for someone who has used a coding agent but not looked under the hood. Part of Willison's Agentic Engineering Patterns guide.

Why it's here
Provenance
seed
Rubric score
9/10
Rubric note
Authoritative practitioner. Durable because harness mechanics outlive model versions. Overlaps with the harness-engineering post on the general framing but is the cleanest intro-level explanation specifically for coding agents.
  • harnesses
  • coding-agents
currentessayadvanced

Effective harnesses for long-running agents

Anthropic · added 2026-07-07

Live link, checked 2026-07-07

(init script, progress file, feature list, git commit), and each coding agent session reads progress, picks one feature, makes incremental progress, and leaves a clean state. The progress-as-JSON-not-markdown and self-verify-before-marking-done rules are concrete and reusable. Essential for any loop that runs across hours or days.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Anthropic engineering, written from building the Claude Agent SDK. The initializer/coding-agent split and the progress-file discipline are not covered elsewhere in the library and are directly actionable for long-running loops.
  • harnesses
  • maintenance
  • context-engineering
currentessayintermediate

Writing effective tools for AI agents

Anthropic · added 2026-07-07

Live link, checked 2026-07-07

How to design tool schemas the model calls reliably, via the Model Context Protocol. The tool-design guidance (names, descriptions, argument shapes, when to split vs combine tools) is the part of harness engineering most people skip and most failures trace back to. Read alongside Osmani's harness post; this is the tools chapter.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Anthropic engineering on a specific, under-covered part of harness design. Concrete schema guidance and MCP framing not duplicated elsewhere in the library. Directly actionable for anyone writing tool definitions.
  • harnesses
  • context-engineering
currentessayadvanced

Rebuilding Devin for Claude Sonnet 4.5: Lessons and Challenges

Cognition · added 2026-07-07

Live link, checked 2026-07-07

test-writing changed the feedback loop shape, subagent delegation got trickier, and meta-agent prompting became viable. A rare production write-up of a harness changing because the model changed — the maintenance/rot problem from the inside.

Why it's here
Provenance
seed
Rubric score
9/10
Rubric note
Cognition engineering, credible. Lower durability because it is anchored to a model release, but the harness-rot-when-the-model-changes lesson is durable and uniquely on-topic for the maintenance pillar.
  • harnesses
  • coding-agents
  • maintenance

loop-engineering

currentessayintro

Loop Engineering

Addy Osmani · added 2026-07-07

Live link, checked 2026-07-07

Defines loop engineering as designing the recurring system that finds work, assigns it, checks it, records state, and decides what runs next — not writing better prompts. The framing the field has converged on and the natural first stop for a reader new to loops. Pairs with the harness-engineering post.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Senior practitioner (Google Chrome) writing from building; coins the "loop engineering" framing the field uses. Model- and framework-independent, directly actionable, and the one entry that frames the meta-practice rather than a single pattern.
  • loop-engineering
currentessayintermediate

Designing agentic loops

Simon Willison · added 2026-07-07

Live link, checked 2026-07-07

Willison's preferred definition — "an LLM agent runs tools in a loop to achieve a goal" treating the wrecking-ball tendency as the thing to engineer against. Short, sharp, and the second-most-cited practitioner framing after Osmani. Read alongside the loop-engineering post.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Willison is a widely trusted practitioner; this post and his "agent is tools in a loop" definition are field-defining. Concrete about bounded loops and stop conditions; distinct from Osmani's loop-engineering framing while covering adjacent ground.
  • loop-engineering
  • guardrails
currentessayintermediate

The Art of Loop Engineering

Sydney Runkle (LangChain) · added 2026-07-07

Live link, checked 2026-07-07

event-driven loop (cron/webhook-triggered runs), and a hill-climbing loop that analyzes traces and rewrites the harness. The fourth loop is the distinctive idea — automate improvement, not just work — and maps directly onto Loopmaster's self-maintenance roster.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
LangChain engineering team; the four-loop stack (esp. the hill-climbing improvement loop) is a distinct and actionable framing not covered elsewhere in the library. Maps onto Loopmaster's own self-maintenance loops, making it uniquely relevant.
  • loop-engineering
  • verification
  • maintenance
  • orchestration
currentessayadvanced

The Agent Loop: State Machines for Production AI

Diesel (bludiesel.ai) · added 2026-07-07

Live link, checked 2026-07-07

Argues the agent loop is an engineering problem (not an AI problem) and state machines error-recovery state, and a dedicated evaluation state that encodes domain knowledge programmatically. State persistence for crash recovery and audit is covered. The production-engineering complement to the patterns essays.

Why it's here
Provenance
seed
Rubric score
9/10
Rubric note
Practitioner blog; authority moderate (1). The state-machine framing with an explicit evaluation state and crash-recovery persistence is a distinct, durable engineering angle not covered by the patterns essays. Directly actionable for production loops.
  • loop-engineering
  • guardrails
  • maintenance
currentessayintermediate

The power of agentic loops — implementing flexbox layout in 3 hours

Colin Eberhardt (Scott Logic) · added 2026-07-07

Live link, checked 2026-07-07

hours, ~800 LOC and ~350 tests, using a browser reference implementation as the feedback mechanism. The empirical lessons — feedback mechanism matters more than the prompt, observe and tune the loop, let it run free once tuned — are the practitioner takeaway. Readable precisely because it is one concrete run, not a survey.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Scott Logic engineering blog; named practitioner. A single concrete run with real numbers and honest caveats. The feedback-mechanism-over-prompt lesson is durable and the case-study shape is unique in the library.
  • loop-engineering
  • verification
  • coding-agents

observability

currenttoolintermediate

Langfuse

Langfuse · added 2026-07-07

Live link, checked 2026-07-07

OSS (MIT) tracing + eval + prompt management; self-host or cloud. The popular open default when you want to own your trace data and run evals on top of production traces. Pairs with a code-first eval framework (DeepEval/Promptfoo) for the standard two-layer

Why it's here
Provenance
seed
Rubric score
9/10
Rubric note
Widely adopted OSS observability+eval tool. Durable as the self-hostable trace layer. Uniqueness is moderate — it overlaps the eval-tools cluster — but the OSS-self-host trace angle is distinct from the CI-gate tools.
  • observability
  • eval-tools
currenttoolintermediate

Arize Phoenix

Arize AI · added 2026-07-07

Live link, checked 2026-07-07

so debug them like an APM. Captures a trace tree of LLM calls, tool calls, retrievals, and sub-agent invocations, and runs evaluators (LLM-as-judge, regex, custom) on top. Self-hostable; the vendor-neutral alternative to LangSmith for teams that want eval on production traces.

Why it's here
Provenance
seed
Rubric score
9/10
Rubric note
Arize AI, OSS. The OTel/OpenInference + trace-as-eval-input framing is a distinct shape from Langfuse's trace+prompt-management; both are durable. Overlaps the observability cluster but the OTel-standard angle is materially different.
  • observability
  • eval-tools

orchestration

currentessayadvanced

How we built our multi-agent research system

Anthropic · added 2026-07-07

Live link, checked 2026-07-07

parallel, with a coordinator synthesizing. Covers the orchestration challenges, context-handoff design, and the trade-offs that made multi-agent worth the overhead. A real production orchestrator-workers case study, not a toy demo.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Anthropic engineering on a shipped production multi-agent system. Distinct from the patterns essay because it is a concrete case study with failure modes and trade-offs. Directly relevant to orchestrator-workers templates.
  • orchestration
  • patterns
  • context-engineering

patterns

currentessayintro

Building Effective Agents

Anthropic · added 2026-07-07

Live link, checked 2026-07-07

The canonical workflow-vs-agent framing and five workflow patterns (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer) that most of the field now references. Argues for starting with the simplest deterministic workflow that works and graduating to a full agent only when control flow is genuinely unknown. Read this before designing any agent system — it is the shared vocabulary.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Anthropic's own engineering team writing from building Claude; defines the five patterns the rest of the field cites. Concrete, framework-agnostic, and the explicit "start with a workflow" rule is directly actionable for loop selection.
  • patterns
  • loop-engineering
  • orchestration
currentrepointermediate

Claude Cookbooks — Agent Patterns

Anthropic · added 2026-07-07

Live link, checked 2026-07-07

Runnable notebooks showing the agent patterns Anthropic ships in production, including tool-use loops, subagents, and multi-step orchestration against Claude. Good for going from "I read Building Effective Agents" to "I can run the loop and watch it." Pairs with the essay; the code makes the abstractions concrete.

Why it's here
Provenance
seed
Rubric score
9/10
Rubric note
Anthropic-maintained runnable examples of the patterns in Building Effective Agents. Slightly less durable than the essay because it tracks API changes, but it is the reference implementation for Claude-specific loop code.
  • patterns
  • coding-agents
currentpaperintermediate

ReAct: Synergizing Reasoning and Acting in Language Models

Yao et al. · added 2026-07-07

Live link, checked 2026-07-07

The paper that formalized the think-act-observe loop — the foundational agent loop almost every framework ships. Shows the interleaving of reasoning traces and tool calls grounded in observations beats pure chain-of-thought on hallucination and error propagation. Foundational reading; everything else builds on this.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Peer-reviewed (ICLR 2023), the origin of the ReAct pattern that underpins all modern tool-using agents. Defines the scratchpad mechanism that makes reasoning auditable. Still cited and still the right place to start.
  • patterns
  • loop-engineering
currentdocintermediate

ReAct — Project Site

Yao et al. · added 2026-07-07

Live link, checked 2026-07-07

Companion site to the ReAct paper with code, prompts, and the benchmark results. Useful if you want to read the paper and then poke at the actual loop rather than reimplement from the abstract. Lighter than the paper; the paper is the primary source.

Why it's here
Provenance
seed
Rubric score
8/10
Rubric note
Author-maintained companion to the ReAct paper. Durable as a reference but largely a pointer to the paper and code; unique value is the runnable artifact, modest beyond the paper itself.
  • patterns
currentpaperintermediate

Reflexion: Language Agents with Verbal Reinforcement Learning

Shinn et al. · added 2026-07-07

Live link, checked 2026-07-07

Adds evaluator feedback and verbal self-reflection stored in an episodic buffer, lifted GPT-4 on HumanEval from 80% to 91% pass@1 — the result that proved a critique-then-revise (tests, compiler, ground truth); without one the critic self-confirms. Essential for the reflection pattern and its limits.

Why it's here
Provenance
seed
Rubric score
10/10
Rubric note
Peer-reviewed (NeurIPS 2023); the Reflexion pattern and the 91% HumanEval result are the standard citation for self-critique loops. Spells out the oracle requirement that stops the pattern from degrading into self-confirming bias.
  • patterns
  • memory
currentessayintro

Lesson 4: agentic design patterns

Addy Osmani · added 2026-07-07

Live link, checked 2026-07-07

A teaching-oriented walk through the four core patterns (ReAct, Reflection, Tool Use, Planning) with when-to-use and trade-offs, plus how to combine them. The intro-level complement to the Anthropic patterns essay — same patterns, gentler gradient. Good first read for someone who found Building Effective Agents dense.

Why it's here
Provenance
seed
Rubric score
9/10
Rubric note
Same authority as the loop/harness posts. Durable intro-level patterns reference. Overlaps the Anthropic patterns essay on substance but the teaching gradient and the combination guidance are distinct.
  • patterns
currentessayintermediate

Agent Workflow Patterns

Arun Baby · added 2026-07-07

Live link, checked 2026-07-07

Four canonical patterns (Reflection, Tool Use/ReAct, Planning, Multi-Agent) with the numbers that made them stick — e.g. Reflexion lifting GPT-4 from 67% to 88% on HumanEval — and a "fifth hidden pattern," LLM-as-a-Judge, as the evaluation layer under all of them. The matched-complexity escalation table (simple → multi-agent) is a handy decision aid for pattern selection.

Why it's here
Provenance
seed
Rubric score
8/10
Rubric note
Practitioner blog; authority moderate. The numbers-backed pattern explanations and the LLM-as-a-Judge-as-fifth-pattern framing are durable and somewhat distinct. Overlaps both Osmani's patterns lesson and the Anthropic essay on the core patterns.
  • patterns
  • eval-tools