Full agentic workflows: when they are worth it, and what to make deterministic

This page looks beyond drafting and similar light uses. When a model writes text and a person checks it, that is not a full agentic workflow by the definition below.

In my experience, full agents stay rare and come with gates, and I use deterministic processes wherever possible. Anthropic’s advice is to start by “finding the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all.” (Building effective agents)

I keep a claim when a public source backs it, vendors and tools included.

What I mean by fully agentic

That is the line the vendors draw: who controls the path.

What a full agentic run looks like

A full agentic run goes start to finish on its own. Vendor guides describe that run in these steps.

1. A task comes in. Agents “begin their work with either a command from, or interactive discussion with, the human user”. (Building effective agents)
2. The model plans. LLMs “dynamically direct their own processes and tool usage”. (Building effective agents)
3. The model chooses tools. OpenAI describes an agent that “dynamically selects the appropriate tools depending on the workflow’s current state”. (A practical guide to building agents)
4. The model acts through tools. OpenAI describes an agent with tools “to gather context and to take actions”. (A practical guide to building agents)
5. It reads the result and loops. Anthropic describes agents as “LLMs using tools based on environmental feedback in a loop”. (Building effective agents)
6. Stop conditions end the loop. Anthropic notes that “it’s also common to include stopping conditions (such as a maximum number of iterations) to maintain control”. (Building effective agents)
7. Human checkpoints stay available. Agents “pause for human feedback at checkpoints or when encountering blockers”. (Building effective agents)

That pause is optional. The model still directs the path, so the run is still fully agentic.

The path is not known in advance.

Assess first, then design

I start with two questions: how often will it be used, and what does a wrong result cost? I run the Assessment framework. Once that call says the work is worth doing and the risk is acceptable, then I use the Solutioning framework. A related article on this site covers the assessment call: The most valuable AI call you’ll make is often “not this one”.

When a full agentic run is justified

A full agentic run can fit simple bug fixes, or routine tasks that have checks built in. Often I can lower the risk by putting gates and checks in place before deployment.

Vendor guides draw a related line.

Can this step be deterministic?

I look at the input and the output. I think about any function I can put between them that is repeatable. I figure out the relationship, and then I write a set of rules.

If the rule set is exhaustive, the step can be deterministic.

If the rules cover only part of the step, I make that part deterministic and I break out the rest as its own step.

When a step should stay agentic

A step should stay agentic if there is no reasonable deterministic rule set. If it really needs thinking and interpretation, and there is no set process, then I can start to argue for an agent.

Lines I hold

Some examples of where I draw the line. Here GxP stands for Good “X” Practice, which the UK Medicines and Healthcare products Regulatory Agency (MHRA) uses in its GxP data integrity guidance as a collective term for good distribution, clinical, laboratory, manufacturing and pharmacovigilance practice.

A fixed step can be tested

Run it against known inputs and you get known outputs. An agent path can differ on every run. Anthropic says: “Agents make dynamic decisions and are non-deterministic between runs, even with identical prompts. This makes debugging harder.” (How we built our multi-agent research system)

A 2024 benchmark, τ-bench, reported that function calling agents built on models like gpt-4o “succeed on <50% of the tasks, and are quite inconsistent (pass^8 <25% in retail)”. (τ-bench) pass^8 is the share of tasks solved on all eight trials. That is a 2024 result, not a forecast for models in use now.

Where the rules stand

The trade-off

On cost to run, Anthropic says “Agentic systems often trade latency and cost for better task performance”. (Building effective agents) In its multi-agent research system, “agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats”. (How we built our multi-agent research system)

Converting costs flexibility, and someone has to own the rules. Google describes a fixed sequence that “can reduce latency and operational costs” compared with model orchestration, but “this efficiency comes at the cost of flexibility”. (Choose a design pattern for your agentic AI system) OpenAI notes rule systems that have become “unwieldy due to extensive and intricate rulesets”. (A practical guide to building agents)

Staying agentic costs predictability, the effort to check the result, and the cost to run. I use the cheapest model that does the job well. Risk decides the model for each step, and how much checking that step gets.

Use the least flexible tool that does the job

I want deterministic processes wherever possible, because they are simpler and easier to validate. The ladder below runs from rung 1, the least flexible, to rung 5, a full agent.

Microsoft puts the warning this way: “Agents can add extra complexity. Some tasks don’t need that complexity.” Where the steps are “clear, repeatable, and follow strict rules, use regular code or nongenerative AI models”. (Business plan for AI agents)

In my experience, most pharma use cases are drafting and analysis tools at rung 3, inside a set process, with a person checking the key outputs. Some add a second AI step that checks the first and scores it against set criteria, but a person still signs off.

The controls on each rung are the ones I require at that rung before I trust a step there.

I rate each step on the ladder, not the whole workflow. One process usually mixes rungs. Say a team handles supplier complaints.

Where people approve

People should have to approve at points where a wrong step would cost a lot of time, or where a wrong step would make later steps fail. What the person checks, and how they judge it, should be set out in a documented process.

OpenAI names two triggers for handing work to a person. (A practical guide to building agents)

1. 1

   "Exceeding failure thresholds"

   OpenAI says to set "limits on agent retries or actions". When an agent goes past them, it should "escalate to human intervention".

2. 2

   "High-risk actions"

   High-risk actions are actions that are "sensitive, irreversible, or have high stakes".

Data foundation

Batch records against lab results

Say a team reconciles batch records against lab results before release. In a full agentic design, the team hands the whole reconciliation to an agent. The agent reads the batch record and the lab results, plans its own path, and writes the disposition, the decision on that batch. It stops when it judges the work complete. No person confirms the findings along the way. A watch on the run is separate from that, and I step in only if it strays.

That is a full agentic run, and I would not use it here. The disposition stays with Quality Assurance (QA).

When a result and the batch record do not match, I handle it in one of three ways. I keep the middle one.

Move up the ladder only when the time saved is worth the extra checking.

This is how I split a step between code and a model.

Which kinds of steps run as code, which use rules with an AI backup, and which run on a model.

Runs as	What it covers
Code:	Extraction from documents with a known layout, format and required-field checks, math and totals, routing, template filling, exact-duplicate removal, and simple yes/no rule checks.
Rules with an AI backup:	Reconciliation.
A model:	Sorting that needs inference, extraction from unstructured text, near-duplicate detection, and judging free text against a policy.

1. 1

   Match lots and check limits

   Rules and queries match lots and check results against limits.

   RuleQuery or script

   Needs: the batch record, the lab results, and the limits for those results.

   Can go wrong: lot identifiers that do not line up, or a result with no limit.

2. 2

   Sort free-text comments

   A limited model step sorts the free-text comments.

   Limited model

   Needs: the free-text comments on the records.

   Can go wrong: a comment placed in the wrong group.

3. 3

   The step left to a read-only agent

   Investigate mismatches

   A read-only agent, with a step limit, investigates mismatches by pulling deviation history and past results, and returns its evidence.

   Constrained agent

   Needs: deviation history and past results.

   Can go wrong: history the agent cannot read, or evidence that does not support the finding.

4. 4

   Confirm before disposition

   QA confirms every finding before disposition.

   QA confirms

   Needs: the findings and the evidence behind them.

   Can go wrong: a finding used before QA has confirmed it.

What this page does not decide

Where a full agent holds up

Competitive intelligence monitoring is usually done by a commercial strategy or market insights team, and it is not a GxP process. An agent searches public trial registries, press releases, and conference abstracts, and drafts a weekly brief with a link for every claim. No one takes a step in that process. The risk is low, so a watch on the run is enough, and I step in only if it strays. A full agentic run holds up here.

---

Want to scope an agentic workflow with me? That is the work I do.