Agentic workflow A multi-step way of using AI where the model iterates instead of answering once.
Short answer. An agentic workflow is a way of using a language model in several passes: it plans, uses tools, checks its own work and revises, instead of producing one answer in a single shot. Andrew Ng named four patterns in March 2024: reflection, tool use, planning and multi-agent collaboration. Anthropic calls both fixed-path workflows and model-directed agents agentic systems, and warns that they often trade latency and cost for better task performance. A tender response shows the idea: find past bids, outline, draft, check against the eligibility list, revise, then hand the result to a person to approve.
01
Four design patterns
Andrew Ng’s March 2024 letter in The Batch proposed four patterns for building agentic systems. They can be combined.
Pattern
What the model does
Office example
Reflection
Examines its own work and looks for ways to improve it
Re-reads a draft client letter for claims it cannot support
Tool use
Calls tools such as search or code execution to gather information or act
Looks up a client’s file number before drafting
Planning
Comes up with, then carries out, a multi-step plan
Outlines a tender response, researches each section, writes the draft
Multi-agent collaboration
Several agents split the work and debate ideas
One agent drafts and another checks every citation
02
Workflow shapes that need no agent
Anthropic separates workflows, where code fixes the path, from agents, where the model directs its own steps. It lists five workflow shapes and advises finding the simplest solution possible, adding complexity only when it demonstrably improves results.
Prompt chaining: split a task into fixed steps, each using the previous output, with checks between them. Example: write an outline, check it meets the criteria, then write the document.
Routing: classify an input and send it to a specialised step. Example: easy questions go to a smaller, cheaper model and hard ones to a larger model.
Parallelisation: run independent parts at the same time, or run one task several times and compare. Example: one call answers a query while another screens it.
Orchestrator and workers: a central model breaks a task into parts it could not predict in advance and hands them to workers.
Evaluator and optimiser: one call writes, another critiques, and the loop repeats until the criteria are met.
03
What iteration buys, and what it costs
In the same letter, Ng reported that GPT-3.5 solved 48.1% of problems on the HumanEval coding benchmark in a single pass and up to 95.1% when wrapped in an agent loop, while GPT-4 scored 67.0% in a single pass. That is a coding result from 2024, not a promise for office documents, but it shows why iteration is popular.
The price is time and money. Anthropic notes that agentic systems often trade latency and cost for better task performance, that errors can compound over many steps, and that it is common to give agents stopping conditions such as a maximum number of iterations. A five-pass workflow makes at least five model calls where a single prompt makes one.
04
Where a person belongs
Add a human gate wherever a mistake would be costly or hard to undo. OpenAI’s guide names two triggers for human intervention: the agent exceeds a limit on retries or actions, or the action is high-risk, such as cancelling orders, authorising large refunds or making payments.
Before anything leaves the building.
Before anything is paid, deleted or shared more widely.
When the workflow has retried more than a set number of times.
When an answer cannot be traced to a source passage.
05
Where ANVE stands
ANVE’s meeting-minutes flow is a small agentic workflow with a human gate: every action item must quote the transcript line it came from, and its owner must be confirmed by a person before it is approved, with each confirmation logged. The agent loop allows up to seven tool steps per request. Agents for tender responses and action follow-up are built with a paid pilot and are not shipped products.
06
Questions people ask
Is an agentic workflow the same as an AI agent?
Not quite. In Anthropic’s terms an agent chooses its own steps, while a workflow follows a path fixed in code. Both are agentic systems, and many useful office systems are workflows.
Does it need a bigger model?
Not necessarily. Ng’s 2024 results showed an older model in an agent loop beating a newer model in a single pass on one coding benchmark. Test on your own task before paying for a larger model or more hardware.
Why are agentic workflows slower?
Each pass is another model call, and tool calls take time too. Anthropic says agentic systems often trade latency and cost for better task performance.
When should I not use one?
When a single well-written prompt with retrieval already works. Anthropic recommends starting with the simplest solution and adding complexity only when it demonstrably improves outcomes.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
