{"licence":{"name":"CC BY-SA 4.0","spdx":"CC-BY-SA-4.0","url":"https://creativecommons.org/licenses/by-sa/4.0/","attribution":"Atlas, a bilingual technical dictionary (https://cmaintz.github.io/tech-atlas/)"},"id":"ai/agentic-workflow","url":{"en":"https://cmaintz.github.io/tech-atlas/en/terms/ai/agentic-workflow/","da":"https://cmaintz.github.io/tech-atlas/da/terms/ai/agentic-workflow/"},"term":{"en":"Agentic workflow","da":"Agentisk arbejdsgang"},"aka":{"en":["LLM workflow"],"da":["LLM-arbejdsgang"]},"domain":["ai"],"cluster":"agents","layer":"agent","status":"emerging","era":2024,"summary":{"en":"A fixed chain of steps written in code, where a language model does some of the steps but does not choose the order.","da":"En fast kæde af trin skrevet i kode, hvor en sprogmodel udfører nogle af trinene, men ikke selv vælger rækkefølgen."},"body":{"formal":{"en":"A system in which large language model calls and tool calls follow paths set in advance by the developer - in a line, in branches chosen by rules, or side by side - rather than the model deciding its own next step as an AI agent does.","da":"Et system, hvor kald til store sprogmodeller og værktøjskald følger veje, som udvikleren har lagt på forhånd - i række, i forgreninger valgt ud fra regler eller side om side - i stedet for at modellen selv bestemmer næste trin, som en AI-agent gør."},"plain":{"en":"Like a relay race on a fixed route - each runner is fast and skilled, but where they run and whom they hand the baton to was settled before the start.","da":"Som et stafetløb på en fast rute - hver løber er hurtig og dygtig, men hvor de løber, og hvem de giver stafetten videre til, blev bestemt før start."},"inPractice":{"en":"At a water utility, each customer email passes a fixed chain - a model sorts it by topic, a prompt written for that topic drafts a reply, code checks the draft against the price rules, and a customer adviser sends it.","da":"Hos et vandværk går hver kundemail gennem en fast kæde - en model sorterer den efter emne, en prompt skrevet til det emne laver et svarudkast, kode tjekker udkastet mod prisreglerne, og en kunderådgiver sender det."},"whyItMatters":{"en":"For most business tasks a fixed path is cheaper, easier to test and harder to lead astray than a free-acting agent, so it is often the better first choice.","da":"Til de fleste forretningsopgaver er en fast vej billigere, lettere at teste og sværere at lede på afveje end en agent, der handler frit, så den er ofte det bedste første valg."}},"deepDive":{"en":"The distinction most practitioners now use comes from Anthropic's \"Building effective agents\" (December 2024): workflows are systems in which LLMs and tools are orchestrated through predefined code paths, whereas agents let the model dynamically direct its own process and tool use. In a workflow the control-flow graph is authored by the developer and is known before execution; the model fills in nodes. The term \"agentic workflow\" is used loosely in industry, sometimes for anything involving an LLM loop, so it pays to ask who owns the next-step decision: code or model.\n\nThe same article names five recurring patterns. Prompt chaining decomposes a task into sequential calls, with programmatic gates between steps (for example, rejecting a draft that fails a length or schema check). Routing classifies the input and dispatches it to a specialised prompt or cheaper model. Parallelisation runs calls concurrently, either by sectioning a task into independent subtasks or by voting, running the same task several times and aggregating, which trades cost for reliability. Orchestrator-workers lets a central LLM decide at run time how to split the work, and evaluator-optimiser loops a generator against a critic until acceptance criteria are met. The last two already hand some control to the model and sit on the boundary with agents.\n\nEngineering a workflow looks like ordinary distributed-systems work. Each step should exchange structured output validated against a schema, so failures are caught at the edge rather than propagating as free text. Steps need timeouts, retries with backoff, idempotency keys for side-effecting tool calls, and checkpointing so a long run can resume; durable-execution engines and graph frameworks such as LangGraph exist largely for this. Because each node has a narrow contract, it can be evaluated in isolation with its own test set, which is the main practical advantage over free-running agents, whose trajectories vary from run to run.\n\nWorkflows also make control placement explicit. A human-in-the-loop approval can be inserted at the one step that sends email or moves money; untrusted content can be confined to steps that have no dangerous tools; and cost and latency are bounded because the number of model calls is fixed or capped. The failure mode is rigidity: inputs the designer did not anticipate fall through the routing logic. A common evolution is to start with a chain, add routing as edge cases appear, and grant agentic autonomy only to the sub-task that genuinely needs open-ended exploration.","da":"Den skelnen, de fleste praktikere bruger i dag, stammer fra Anthropics \"Building effective agents\" (december 2024): workflows er systemer, hvor LLM'er og værktøjer orkestreres gennem foruddefinerede kodestier, mens agenter lader modellen selv styre sin proces og sin brug af værktøjer dynamisk. I en agentisk arbejdsgang er kontrolflowgrafen skrevet af udvikleren og kendt før afvikling; modellen udfylder knuderne. Betegnelsen bruges løst i branchen, nogle gange om alt, der indeholder en LLM-løkke, så det betaler sig at spørge, hvem der ejer beslutningen om næste trin: koden eller modellen.\n\nSamme artikel navngiver fem tilbagevendende mønstre. Prompt chaining deler en opgave op i sekventielle kald med programmatiske kontrolpunkter imellem (fx afvises et udkast, der ikke består et tjek af længde eller skema). Routing klassificerer inputtet og sender det videre til en specialiseret prompt eller en billigere model. Parallelisering kører kald samtidig, enten ved sectioning, hvor opgaven deles i uafhængige delopgaver, eller ved voting, hvor samme opgave køres flere gange og resultaterne samles, hvilket bytter omkostning for pålidelighed. Orchestrator-workers lader en central LLM beslutte under kørsel, hvordan arbejdet deles, og evaluator-optimizer kører en generator i løkke mod en kritiker, indtil acceptkriterierne er opfyldt. De to sidste overlader allerede noget kontrol til modellen og ligger på grænsen til agenter.\n\nAt bygge en arbejdsgang ligner almindeligt arbejde med distribuerede systemer. Hvert trin bør udveksle struktureret output, der valideres mod et skema, så fejl fanges ved grænsen i stedet for at brede sig som fritekst. Trin skal have timeouts, genforsøg med backoff, idempotensnøgler til værktøjskald med sideeffekter og checkpoints, så en lang kørsel kan genoptages; motorer til durable execution og grafframeworks som LangGraph findes i høj grad af den grund. Fordi hver knude har en snæver kontrakt, kan den evalueres isoleret med sit eget testsæt, og det er den største praktiske fordel frem for frit kørende agenter, hvis forløb varierer fra kørsel til kørsel.\n\nArbejdsgange gør også placeringen af kontroller eksplicit. En menneske i løkken-godkendelse kan indsættes netop ved det trin, der sender e-mail eller flytter penge; upålideligt indhold kan holdes i trin uden farlige værktøjer; og omkostning og latenstid er begrænset, fordi antallet af modelkald er fast eller har et loft. Fejlmåden er stivhed: input, som designeren ikke forudså, falder igennem routinglogikken. En typisk udvikling er at starte med en kæde, tilføje routing, efterhånden som særtilfælde dukker op, og kun give agentisk selvstændighed til den delopgave, der reelt kræver åben udforskning."},"edges":[{"type":"requires","to":"ai/large-language-model","confidence":"high","strength":"normal"},{"type":"requires","to":"ai/prompt","confidence":"high","strength":"normal"},{"type":"contrasts-with","to":"ai/ai-agent","why":{"en":"In a workflow the developer's code decides the order of steps; in an AI agent the model decides for itself as it goes.","da":"I en arbejdsgang bestemmer udviklerens kode rækkefølgen af trin; i en AI-agent bestemmer modellen selv undervejs."},"confidence":"high","strength":"primary"},{"type":"used-with","to":"ai/structured-output","why":{"en":"Steps pass their results to the next step in a fixed shape, so the code can check and route them reliably.","da":"Trinene giver deres resultater videre til næste trin i en fast form, så koden kan tjekke og dirigere dem pålideligt."},"confidence":"high","strength":"normal"},{"type":"used-with","to":"ai/human-in-the-loop","why":{"en":"Because the steps are known in advance, it is easy to place a person's approval at exactly the risky step.","da":"Fordi trinene er kendt på forhånd, er det let at lade en person godkende netop det risikable trin."},"confidence":"medium","strength":"normal"},{"type":"used-with","to":"ai/multi-agent-system","confidence":"high","strength":"normal"}],"depth":4,"sources":[{"title":"Anthropic (2024), Building effective agents","tier":"reference","publisher":"Anthropic"}],"draft":true}