Agentic tooling (umbrella)
The tools for building and running LLM agents — everything wrapped around a frontier model to turn it from a text generator into a system that does work. This wiki's umbrella concept. Its premise, drawn straight from the corpus: the model writes the code; the value has moved to everything around it. That "everything" decomposes into the layers the synthesis tracks:
- Capability as portable markdown — domain skills bundled as files an agent loads at runtime (agent-skills cross-wiki), hardening into an open cross-vendor standard (agentskills-spec) alongside its sibling conventions agents-md and model-context-protocol (skills × MCP = agency).
- Structure around the model — the agentic-coding-harness and its composable agent-middleware: phase gates, deterministic tools, and confidence grading that let structure substitute for raw capability.
- Execution & authoring — agent-orchestration (orchestrator → parallel-subagent fan-out) at run time, spec-driven-development at authoring time, and the shift from prompt to loop — goal + stop-condition + feedback signal — where the bottleneck becomes verification, not generation.
- Durability & self-improvement — durable-agents that pause/resume across crashes, self-improving-agents that author and refine their own skills, over an agent-memory state substrate.
- Bounds — agent-guardrails: autonomy capped by reversibility, human checkpoints on the irreversible — the counterweight to the autonomy push above.
What counts as an "agent"
Anthropic draws the line this spoke builds on (building-effective-agents): a workflow is "systems where LLMs and tools are orchestrated through predefined code paths," whereas an agent is "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks." The agent steers its own loop; the workflow runs a fixed one. And the agent's competence rests on a feedback detail — "it's crucial for the agents to gain 'ground truth' from the environment at each step" — the same point loop-engineering makes about verification.
The through-line
Across every layer the same bet recurs: structure, not model size, is the lever. Skills, harnesses, loops, durability, and guardrails are all reliability engineering around a capability you don't control — which is why the open question the spoke keeps returning to is whether structure really substitutes for capability, and how far. Worth holding against the spoke's maximalism: the first-party advice (building-effective-agents) is restraint — add complexity "only when it demonstrably improves outcomes." See synthesis.
The layers, argued as one thing
hitchhikers-guide-agentic-ai (Haggai Roitman, arXiv, June 2026) is the corpus's only book-length attempt to hold this whole umbrella at once, and its thesis is about the practitioner rather than the architecture: building good agentic systems requires understanding every layer of the pipeline, not one of them. That is a claim this wiki's own structure keeps testing. The spoke pages the layers separately — skills, harness, framework, orchestration, evals — because each has its own sources and its own failures; the guide's argument is that anyone who reads only one of those pages has learned the wrong lesson.
Related
hitchhikers-guide-agentic-ai · building-effective-agents · agentic-coding-harness · agent-middleware · agent-orchestration · loop-engineering · durable-agents · self-improving-agents · agent-guardrails · agentskills-spec · agents-md · agent-memory · synthesis