On this article, you’ll discover ways to distinguish agentic workflows from autonomous brokers by specializing in who owns management stream — a human writing code upfront, or a mannequin reasoning at runtime.
Matters we’ll cowl embody:
- Why the true axis separating these programs is predictability versus autonomy, not whether or not an LLM is concerned.
- How deterministic workflows, orchestrated workflows, reactive brokers, and autonomous multi-agent programs differ, with runnable code that makes the control-flow distinction concrete.
- Why workflows, not totally autonomous brokers, dominate manufacturing in the present day, and why hybrid architectures are the sample that holds up.
Introduction
Deloitte projects that by 2027, as much as 50% of firms utilizing generative AI can have launched agentic AI pilots or proofs of idea. That’s a wave of adoption sufficiently big that the phrase “agentic” has began masking nearly something with an LLM name in it, from a hard and fast five-step pipeline the place step three occurs to name GPT for a abstract to a completely self-directing system that plans its personal path with no script in any respect.
These are usually not the identical factor. Treating them as interchangeable results in certainly one of two errors: over-engineering a easy, well-understood process with pointless autonomy, or under-engineering a genuinely open-ended drawback by forcing it right into a inflexible pipeline that breaks the second actuality deviates from the plan.
Anthropic attracts the foundational line of their broadly cited “Constructing Efficient Brokers” piece: workflows are programs the place LLMs and instruments are orchestrated via predefined code paths. Brokers are programs the place LLMs dynamically direct their very own course of and power utilization, sustaining management over how they accomplish a process. Every thing on this article is detailed beneath that one distinction.
This piece maps the complete spectrum of deterministic workflows, orchestrated programs, reactive single brokers, and autonomous multi-agent programs, with code at every stage that makes the control-flow distinction concrete fairly than summary. The code right here illustrates structure, not a deployable system; the purpose of every snippet is to indicate who decides what occurs subsequent, to not ship a function.
The Actual Axis Isn’t “AI vs. No AI”: It’s Predictability vs. Autonomy
Earlier than evaluating architectures, it’s value changing the incorrect query. The query isn’t “does this method use an LLM.” Nearly all the things does now. The 2 questions that truly matter, borrowing a framing that’s gained actual traction in structure circles, are: does this course of must be repeatable, auditable, and explainable step-by-step? And: is the right path even identified upfront, or does the system want to find it at runtime?
A system can lean closely on an LLM and nonetheless be totally deterministic in construction — a hard and fast pipeline the place one step occurs to name a mannequin for textual content technology, however the subsequent step is hardcoded no matter what comes again. A system will also be “agentic” with little or no actual autonomy: a tightly scripted loop with solely two allowed actions and a tough step restrict. The presence of an LLM name just isn’t the sign. Possession of management stream is.
Google Cloud’s personal design-pattern documentation attracts this actual line operationally: deterministic workflows embody duties with a clearly outlined path identified upfront, the place the steps don’t change a lot from one run to the following. Workflows that require dynamic orchestration contain issues the place the agent should decide one of the simplest ways to proceed, and not using a predefined script. That’s the spectrum this text walks via, one stage at a time.
Deterministic Workflows
That is the baseline. A deterministic workflow has a identified sequence of steps determined at design time, by a human, in code. An LLM can sit inside any step — producing textual content, classifying enter, drafting a abstract — nevertheless it doesn’t select what occurs after its personal step runs. The orchestrating code does that, no matter what the mannequin returns.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 |
# deterministic_pipeline.py # Conditions: none past Python’s customary library # Run: python deterministic_pipeline.py
def mock_llm_classify(textual content: str) -> str: “”“ Mock LLM name — stands in for an actual API name to maintain this instance runnable with out an API key. The purpose is structural: no matter this returns, the NEXT perform that runs is already determined under. ““” if “refund” in textual content.decrease() or “cost” in textual content.decrease(): return “billing” return “common”
def extract(raw_input: str) -> str: “”“Step 1 — at all times runs, at all times results in step 2. No branching right here.”“” return raw_input.strip()
def classify(cleaned_text: str) -> str: “”“ Step 2 — calls an LLM to supply a label, however the label has no impact on which perform runs subsequent. That is the deterministic half: the mannequin fills in a bit of knowledge, it does not affect the route. ““” label = mock_llm_classify(cleaned_text) print(f” [classify] LLM returned label=”{label}” (informational solely)”) return cleaned_text
def summarize(cleaned_text: str) -> str: “”“Step 3 — at all times runs after step 2, whatever the label from step 2.”“” return f“Abstract: {cleaned_text[:40]}…”
def notify(abstract: str) -> str: “”“Step 4 — at all times runs final. The trail is mounted at design time.”“” return f“Notification despatched: {abstract}”
def run_deterministic_pipeline(raw_input: str) -> str: “”“ The management stream right here is written completely by a human, upfront. Each run takes the equivalent path: extract -> classify -> summarize -> notify. The LLM name inside classify() produces a label, however that label is rarely used to determine what perform runs subsequent — it is information flowing via a hard and fast pipe. ““” step1 = extract(raw_input) step2 = classify(step1) step3 = summarize(step2) step4 = notify(step3) return step4
if __name__ == “__main__”: # Two inputs that the LLM would classify utterly otherwise result_1 = run_deterministic_pipeline(“I need a refund for my final cost”) result_2 = run_deterministic_pipeline(“What are what you are promoting hours?”)
print(f“nResult 1: {result_1}”) print(f“Consequence 2: {result_2}”) |
The best way to run: python deterministic_pipeline.py, no dependencies required.
Output:
|
[classify] LLM returned label=‘billing’ (informational solely) [classify] LLM returned label=‘common’ (informational solely)
Consequence 1: Notification despatched: Abstract: I need a refund for my final cost... Consequence 2: Notification despatched: Abstract: What are your enterprise hours?... |
Discover what occurred: the mock LLM labeled the 2 inputs utterly otherwise, billing versus common, and it made zero distinction to the trail both enter took. Each went via the very same 4 features in the identical order. That’s the complete definition of deterministic: the route is mounted, even when an LLM is doing actual work inside one of many steps.
Orchestrated Workflows
That is the center floor that will get mislabeled most frequently as “agentic,” and it’s value slowing down right here as a result of it’s the road most individuals truly cross after they begin utilizing that phrase loosely.
An orchestrated workflow nonetheless has a graph of doable paths outlined completely upfront, however which path will get taken now is determined by a runtime resolution, incessantly made by an LLM name. That is nonetheless a workflow. Each department that may very well be taken was anticipated and written into code by a human earlier than the system ever ran. The LLM picks a department off a menu another person wrote. It doesn’t invent a brand new merchandise on that menu.
That is exactly the “dynamic orchestration” class Google Cloud separates from real brokers — the system must plan and route, however inside a construction {that a} human nonetheless totally designed.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 |
# orchestrated_pipeline.py # Conditions: none past Python’s customary library # Run: python orchestrated_pipeline.py
def mock_llm_classify(textual content: str) -> str: “”“Mock LLM classification name.”“” text_lower = textual content.decrease() if “refund” in text_lower or “cost” in text_lower: return “billing” if “crash” in text_lower or “error” in text_lower or “bug” in text_lower: return “technical” return “common”
def extract(raw_input: str) -> str: return raw_input.strip()
# Three pre-defined downstream handlers. A human wrote all three of those # upfront. The LLM doesn’t invent a fourth path — it could possibly solely choose # amongst branches that exist already on this code. def handle_billing(textual content: str) -> str: return f“[BILLING TEAM] Routed: {textual content[:50]}”
def handle_technical(textual content: str) -> str: return f“[TECH SUPPORT] Routed: {textual content[:50]}”
def handle_general(textual content: str) -> str: return f“[GENERAL QUEUE] Routed: {textual content[:50]}”
# The department map IS the complete resolution area. Each key right here was written # by a human forward of time. The LLM’s job is to choose a key — not outline one. ROUTE_MAP = { “billing”: handle_billing, “technical”: handle_technical, “common”: handle_general, }
def run_orchestrated_pipeline(raw_input: str) -> str: “”“ Nonetheless a workflow, not an agent: each doable path was anticipated and coded by a human forward of time, sitting in ROUTE_MAP. The LLM name decides WHICH pre-built department executes for this particular enter, however it can not invent a department that is not already a key in ROUTE_MAP. ““” cleaned = extract(raw_input) label = mock_llm_classify(cleaned) print(f” [route] LLM labeled as ‘{label}’ -> dispatching to handle_{label}()”)
handler = ROUTE_MAP.get(label, handle_general) return handler(cleaned)
if __name__ == “__main__”: test_inputs = [ “I was charged twice for my refund request”, “The app keeps crashing with an error on startup”, “What are your business hours?”, ] for inp in test_inputs: consequence = run_orchestrated_pipeline(inp) print(f” Consequence: {consequence}n”) |
The best way to run: python orchestrated_pipeline.py, no dependencies required.
Output:
|
[route] LLM labeled as ‘billing’ -> dispatching to handle_billing() Consequence: [BILLING TEAM] Routed: I was charged twice for my refund request
[route] LLM labeled as ‘technical’ -> dispatching to handle_technical() Consequence: [TECH SUPPORT] Routed: The app retains crashing with an error on startup
[route] LLM labeled as ‘common’ -> dispatching to handle_general() Consequence: [GENERAL QUEUE] Routed: What are your enterprise hours? |
Three totally different inputs took three totally different paths this time — that’s new in comparison with the earlier part. However have a look at ROUTE_MAP: each doable vacation spot was already written into the code earlier than any of those inputs arrived. The LLM exercised judgment about which key to make use of. It by no means had the choice to create a key that wasn’t there. That distinction — a hard and fast set of doable paths versus a path that will get invented at runtime — is strictly the place the following part picks up.
Reactive Brokers: The ReAct Loop and a Genuinely Open Path
That is the place actual autonomy begins. The ReAct sample — Reasoning plus Performing, launched by Yao et al. in 2022 — lets the mannequin itself determine, at every step, what motion to take subsequent based mostly on what it noticed from the earlier motion. There isn’t any pre-written department masking each case. The agent operates in an iterative loop of thought, action, and observation till an exit situation is met, and the sequence itself — what number of steps, in what order, and which instruments — just isn’t knowable upfront. Solely the accessible actions are mounted; the trail via them just isn’t.
That is the architectural threshold the earlier two sections had been constructing towards. Within the orchestrated workflow, a human wrote each doable department into ROUTE_MAP earlier than the system ran. Right here, the mannequin decides each the trail and the sequence size at runtime, despite the fact that the toolset itself remains to be mounted.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 |
# react_loop.py # Conditions: none past Python’s customary library # Run: python react_loop.py
def search_knowledge_base(question: str) -> str: “”“A instrument the agent can name. Whether or not and when it will get known as just isn’t determined right here — it is determined by the mannequin, at runtime.”“” mock_kb = { “refund coverage”: “Refunds can be found inside 30 days of buy.”, “delivery time”: “Normal delivery takes 5-7 enterprise days.”, } for key, worth in mock_kb.objects(): if key in question.decrease(): return worth return “No matching data present in information base.”
def escalate_to_human(cause: str) -> str: “”“A second instrument the agent can name — once more, the choice to name this as a substitute of the search instrument is made by the mannequin, not by this code.”“” return f“Escalated to human agent. Purpose: {cause}”
AVAILABLE_TOOLS = { “search_knowledge_base”: search_knowledge_base, “escalate_to_human”: escalate_to_human, }
def mock_llm_decide_next_step(observations: checklist[str], user_query: str) -> dict: “”“ Mock LLM name standing in for the REASONING step of ReAct. In an actual system, that is an precise mannequin name that reads the complete Thought -> Motion -> Statement historical past and decides what occurs subsequent. Critically: this perform — not the calling loop under — decides which instrument to name and when to cease. There isn’t any “if question incorporates X, name Y“ department written wherever in run_react_loop(). The choice is made recent, from collected context, on each single iteration. ““” if not observations: return { “thought”: “I must search for the coverage earlier than I can reply.”, “motion”: “search_knowledge_base”, “action_input”: user_query, } last_observation = observations[–1] if “No matching data” in last_observation: # This department was by no means written by a human upfront — the mannequin # determined, based mostly on what it simply noticed, that escalation was wanted. return { “thought”: “The information base has no reply. I ought to escalate this.”, “motion”: “escalate_to_human”, “action_input”: “No KB match for: “ + user_query, } return { “thought”: “I discovered the reply. Process full.”, “motion”: “end”, “action_input”: last_observation, }
def run_react_loop(user_query: str, max_steps: int = 5) -> str: “”“ Thought -> Motion -> Statement, repeated till the mannequin itself decides to cease. Evaluate this instantly towards run_orchestrated_pipeline() within the earlier part: there isn’t a ROUTE_MAP right here. There isn’t any human-written department saying “if X occurred, name Y.” Each resolution about what occurs subsequent is made by the mannequin, at runtime, based mostly on what it has noticed thus far. ““” observations: checklist[str] = []
for step in vary(max_steps): resolution = mock_llm_decide_next_step(observations, user_query) print(f” Step {step + 1} — Thought: {resolution[‘thought’]}”)
if resolution[“action”] == “end”: return f“Closing reply: {resolution[‘action_input’]}”
tool_fn = AVAILABLE_TOOLS.get(resolution[“action”]) if tool_fn is None: return f“Error: mannequin requested unknown instrument ‘{resolution[‘action’]}'”
statement = tool_fn(resolution[“action_input”]) print(f” Step {step + 1} — Motion: {resolution[‘action’]}({resolution[‘action_input’]!r})”) print(f” Step {step + 1} — Statement: {statement}n”) observations.append(statement)
return “Max steps reached with out decision.”
if __name__ == “__main__”: print(“=== Question A: answerable from the information base ===”) result_a = run_react_loop(“What’s the refund coverage?”) print(f“Consequence: {result_a}n”)
print(“=== Question B: not within the information base, ought to set off escalation ===”) result_b = run_react_loop(“Are you able to course of my worldwide tax refund in crypto?”) print(f“Consequence: {result_b}”) |
The best way to run: python react_loop.py, no dependencies required.
Output:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 |
=== Question A: answerable from the information base === Step 1 — Thought: I want to look up the coverage earlier than I can reply. Step 1 — Motion: search_knowledge_base(‘What’s the refund coverage?’) Step 1 — Statement: Refunds are accessible inside 30 days of buy.
Step 2 — Thought: I discovered the reply. Process full. Consequence: Closing reply: Refunds are accessible inside 30 days of buy.
=== Question B: not in the information base, ought to set off escalation === Step 1 — Thought: I want to look up the coverage earlier than I can reply. Step 1 — Motion: search_knowledge_base(‘Are you able to course of my worldwide tax refund in crypto?’) Step 1 — Statement: No matching data discovered in information base.
Step 2 — Thought: The information base has no reply. I ought to escalate this. Step 2 — Motion: escalate_to_human(‘No KB match for: Are you able to course of my worldwide tax refund in crypto?’) Step 2 — Statement: Escalated to human agent. Purpose: No KB match for: Can you course of my worldwide tax refund in crypto?
Step 3 — Thought: I discovered the reply. Process full. Consequence: Closing reply: Escalated to human agent. Purpose: No KB match for: Can you course of my worldwide tax refund in crypto? |
Have a look at what differs between the 2 runs: question A completed in two steps, question B took three, and question B took an motion — escalation — that was by no means hardcoded as “what occurs when refund queries point out crypto.” The identical loop, the identical code, produced two genuinely totally different step counts and sequences as a result of the mannequin determined the trail at runtime based mostly on what it noticed. That’s the precise, concrete which means of “no predefined code path” — not a slogan, however a measurable distinction in what number of steps had been run and what they had been.
Manufacturing implementations of this sample usually wrap the collected thought/statement historical past in a “scratchpad” and summarize instrument outputs earlier than feeding them again into the loop, since dumping uncooked error logs or massive API responses again into context tends to confuse the following reasoning step fairly than assist it.
Autonomous Multi-Agent Programs
The far finish of the spectrum builds instantly on the ReAct loop above, simply nested. In a multi-agent setup, an orchestrator runs its personal ReAct loop, the place a few of its accessible “actions” are calls to different brokers, every of which runs its personal full ReAct loop inside. The orchestrator causes about what to delegate, delegates it, observes the consequence, and continues — precisely just like the single-agent loop within the earlier part, besides a few of its “instruments” are complete brokers fairly than easy features.
Image the AVAILABLE_TOOLS dictionary from the earlier instance, besides as a substitute of search_knowledge_base and escalate_to_human, the entries are research_agent, finance_agent, and coding_agent — and calling certainly one of them doesn’t return a easy string; it kicks off that sub-agent’s personal impartial Thought-Motion-Statement loop, which could run for a number of steps earlier than returning something to the orchestrator. No person wrote down upfront which sub-agent will get known as, in what order, or what number of instances any of them run.
Google Cloud’s documentation labels essentially the most excessive model of this the “swarm” sample — a collaborative staff of brokers with no central orchestrator in any respect, able to producing exceptionally high-quality, artistic options exactly as a result of nothing is constraining how they work together. That very same lack of construction can be the chance: and not using a human-designed certain on the interplay, a swarm can fall into unproductive loops or just fail to converge, and the price of operating many brokers via many turns compounds shortly.
That is the purpose on the spectrum the place the predictability axis from the primary part swings hardest within the different route. A deterministic pipeline provides you a similar output construction each time, by development. A swarm of autonomous brokers provides you the flexibleness to deal with an issue no one anticipated, at the price of having the ability to predict, upfront, what it can do or how lengthy it can take to do it.
Why This Distinction Truly Issues in Manufacturing
This isn’t an instructional distinction. It has a direct, measurable impact on what groups truly ship. Regardless of the amount of hype round autonomous brokers, AI workflows — not totally autonomous brokers — received the manufacturing battle in 2025: workflows stay the dominant sample behind profitable generative AI deployments, whereas totally autonomous multi-agent programs are nonetheless largely exploratory exterior of slim domains.
The rationale maps instantly again to the predictability axis from the beginning of this text. Agentic programs are non-deterministic by nature; equivalent inputs can produce totally different outputs throughout separate runs, which is a critical legal responsibility in regulated, auditable, or in any other case high-stakes processes. If a course of should be explainable step-by-step to a compliance staff or a regulator, that’s not agent territory by default; it wants guardrails and human-in-the-loop checkpoints layered on prime earlier than it may be trusted with actual penalties.
The sample that’s truly rising in mature programs is hybrid, not a pick-one resolution. A better-level agent units objectives and orchestrates the general process, whereas crucial, well-understood computations nonetheless run inside deterministic modules {that a} human has totally specified. A medical diagnostics system, for instance, would possibly use an agent to interpret ambiguous signs and determine which checks to order — real autonomy, as a result of the appropriate sequence of checks isn’t knowable upfront — whereas every check itself runs via a validated, deterministic pipeline, as a result of that a part of the issue has a identified appropriate path and no cause to introduce variability into it.
Conclusion
“Agentic workflow” and “autonomous agent” describe two ends of 1 spectrum, not two competing applied sciences, and the 4 phases walked via right here — deterministic, orchestrated, reactive, and autonomous multi-agent — aren’t a rating from worse to higher. They’re totally different solutions to the identical query: who decides what occurs subsequent, and was that call made by a human writing code upfront, or by a mannequin reasoning at runtime?
Deterministic workflows offer you auditability and repeatability by development; the identical enter takes the identical path each time, full cease. Reactive and multi-agent programs surrender that assure in change for the power to deal with issues whose form genuinely can’t be anticipated forward of time. Neither property is free, and neither structure is appropriate by default.
The programs that maintain up nicely in manufacturing don’t choose one excessive of this spectrum and apply it in every single place. They place each bit of the issue on the level on the spectrum that piece truly requires — a hard and fast construction wherever a identified appropriate path exists and repeatability issues, with actual autonomy reserved for the components of the issue that haven’t any predefined appropriate path to comply with within the first place.
