Topics
18 categories, each with a maintained overview
Tools & Integrations
Ready-made capabilities an agent can call: search, code execution, file handling and service integrations.
Agent Frameworks
Libraries that give an agent its control loop: how it plans, calls tools, keeps state and recovers from failure. The choice here shapes everything downstream, so weigh debuggability and lock-in alongside how fast the first demo comes together.
Coding Agents
Agents that read, write and run code against a real repository. Benchmark scores are a starting point, not an answer: sandboxing, token cost on large codebases and how reviewable the output is matter more day to day.
Observability & Evals
Seeing what an agent actually did — traces, costs, latencies — and turning those runs into evaluation sets. Without this, an agent in production is a black box you cannot improve.
Multi-Agent Orchestration
Coordinating several agents or steps: routing, delegation, parallel execution and merging results. Worth reaching for only once a single agent has demonstrably hit its limit.
Protocols & Interop
Shared interfaces that let agents, tools and models from different vendors work together — MCP, A2A and the OpenAI-compatible surface most runtimes now expose.
RAG & Retrieval
Retrieval pipelines that ground an agent in your own documents: chunking, embedding, reranking and citation. Retrieval quality, not model choice, is usually what decides whether answers are trustworthy.
Guardrails & Security
Constraining what an agent may say and do: input and output filtering, prompt-injection defence, and policy enforcement around tool calls.
Computer & Browser Use
Driving a browser or desktop on the user's behalf. The dividing line is whether the agent reasons over structured elements or raw pixels — the former is far more reliable wherever a usable DOM exists.
Data & Ingestion
Getting real-world content into a form an agent can use — crawling, scraping, parsing PDFs and normalising documents.
Workflow & Low-code
Visual and low-code builders that place agents inside broader automation. Check the licence carefully: several popular options in this category are source-available rather than open source.
Benchmarks & Datasets
Standard tasks and datasets for comparing agents. Useful for direction, but read what a benchmark actually measures before treating a leaderboard position as a selection criterion.
Memory & Context
Persisting what an agent learns between sessions, instead of replaying an ever-growing transcript. The hard part is deciding what deserves to be remembered — verify extraction quality on your own data.
Vector & Storage
Storage engines for embeddings. Beyond raw search speed, look at metadata filtering, hybrid search and what happens to recall when the collection grows past what fits in memory.
Model Serving
Running models on your own hardware behind an OpenAI-compatible endpoint. Agents are unusually chatty, so throughput and concurrency matter more here than single-request latency.
Sandboxes & Runtimes
Isolated environments where agent-generated code can run without endangering the host. A prerequisite, not an optional extra, once an agent executes anything it wrote itself.
Prompt & Optimization
Treating prompts and agent programs as something to be measured and improved automatically, rather than hand-tuned indefinitely.
Voice Agents
Real-time speech agents. Latency budgets are brutal — interruption handling and turn-taking usually decide whether the result feels usable.