A year ago, our typical "AI automation" project meant a hand-coded workflow with a few LLM calls bolted in. Today it's increasingly an agent: something we hand goals, tools, and a budget to and let figure out the rest. Here's what changes, what breaks, and how we now design for failure from day one.
The short answer: workflow automation follows a graph you wrote in advance; an AI agent is given a goal, tools and a budget and decides its own next step. Agents trade brittleness for variability, so production agent systems need budget ceilings, human approval on irreversible actions, and eval suites that test failure modes. That is the design discipline AGI Software Solutions applies to every agent build.
One framing note before the detail. Buyers usually come to this looking for workflow automation services and discover partway through that the interesting question is not which tool to use, it is which parts of a process should follow a fixed graph and which need something that can decide. Get that split right and the build is straightforward; get it wrong and you either ship a brittle graph that breaks weekly or an expensive agent doing work a script handled fine. Our AI automation services page covers how we scope that split, and RPA vs AI agents is the decision table version.
The shift
Workflow automation is a graph. You write the nodes, you write the edges, and the system follows them. If something unexpected happens, the workflow halts or branches to an error handler you wrote.
An agent works differently. You give it a goal, a set of tools, and a budget, and it decides what to do next. When it hits an unexpected state, it tries to recover on its own, sometimes well, sometimes hilariously badly.
What changes
- Less brittleness, more variability. Workflows fail clean. Agents fail weird.
- Latency goes up. An agent that has to plan, call tools, and reflect takes seconds, not milliseconds.
- Cost is unpredictable. A workflow has a fixed call count. An agent might make 3 calls or 30.
- Observability becomes critical. When something goes wrong, you need to replay the agent's reasoning.
How we design for failure
The single most important design principle: budget ceilings. Time budgets, call-count budgets, and dollar budgets. Hit any limit → agent stops → human is paged.
Beyond that, three patterns we've found essential:
- Reflection checkpoints. After every K tool calls, the agent summarizes what it's done, why, and what's left. We log this. If the summary stops making sense, we kill the run.
- Reversibility ratchets. Group tools by reversibility. Cheap-and-reversible (search, read) → unlimited. Expensive-or-irreversible (send email, charge card) → human approval required.
- Eval suites that include failure modes. Don't just test that the agent solves the happy path. Test that it gives up cleanly when the path doesn't exist.
Where this is going
We think the next 12 months are about multi-agent systems: small teams of specialized agents that hand work off to each other, with a coordinator watching budgets and quality. We're already shipping early versions of this for two clients, and we'll write more about it once the pattern settles.
Frequently asked questions
What is the difference between an AI agent and workflow automation?
A workflow is a graph you write: nodes, edges and error handlers authored in advance. An agent gets a goal, tools and a budget and decides its own next step, including how to recover. Workflows fail clean; agents fail weird — which is why agents need budgets, checkpoints and evals.
When should a business use an agent instead of a scripted workflow?
Scripted workflow when the process is stable and enumerable. Agent when the path varies per case and enumerating every branch is impractical. Many production systems combine both: a workflow spine with agent steps where flexibility pays.
How do you stop an agent running away with costs or bad actions?
Budget ceilings — time, call-count and dollars; hit any limit and the agent stops and a human is paged. Then reversibility ratchets: reads run free, irreversible actions need human approval. And eval suites that test the agent gives up cleanly when the path doesn't exist.
What is a multi-agent AI system?
A small team of specialized agents handing work off to each other, with a coordinator watching budgets and quality. Worth it when the work is too broad for one agent's context or toolset; the price is more orchestration and observability engineering.