AI agents · workflow practice

If you can write the runbook,
we can build the agent.

Workflow agents that actually finish the job: read the input, call the tools, handle the failures and log every step. With approval gates where a human should stay in the loop.

Describe your workflow All AI agent work

What it is

What workflow automation with agents means

Most back-office work is a runbook someone follows by hand: pull the document, check it against a system, decide, write the result somewhere, tell a person. Each step is small. The cost is that a human has to hold the whole sequence.

An agent runs that sequence. The engineering is in the unglamorous parts: what happens when the third API times out, what the agent is allowed to do without asking, how a half-finished run resumes, and how you audit a decision three months later.

What we build

Agents with tools, memory and brakes.

Durable execution, explicit permissions and a trace for every run.

Document workflows

Intake, extraction, validation against your systems, and a structured result. Invoices, KYC packs, claims, contracts.

Schema-validated outputConfidence thresholdsException queue

Multi-step pipelines

Several agents with distinct jobs, coordinated with durable execution so a failure resumes instead of restarting.

Durable runsRetries and backoffResume from failure

Tool and system access

Your CRM, ERP, ticketing or internal APIs, with scoped credentials and an allow-list of what the agent may call.

Scoped credentialsTool allow-listRate limiting

Human approval gates

The agent pauses where the stakes justify it, routes to a person, and continues with the decision recorded.

Approval routingDecision logTimeout policy

Audit trails

Every input, tool call, model response and output stored and replayable, so you can answer why it did that.

Replayable runsPer-step logsExport for audit

Kill switch

One control that stops every run mid-flight, plus per-workflow pause. Nothing keeps executing while you investigate.

Global stopPer-workflow pauseSafe rollback

Built on

The stack we reach for.

Boring where it should be boring. The parts your next hire already knows how to run.

ClaudeGPTTemporalLangGraphPostgresModalLangfuseDockerOpenTelemetry

The process

How an automation build runs.

We start from your runbook and end with something your team can watch running.

Runbook review

Tool inventory

Happy path

Failure paths

Approvals and audit

Handover

Common questions

Automation questions we get first.

Not answered here? Email us. A real reply from an engineer, no form.

Is an agent the right answer for our process?

Sometimes not. If the process never varies, a script is cheaper and more predictable. Agents earn their keep where the input is messy or the next step depends on judgement. We say so on the first call rather than selling you one.

What happens when a tool call fails?

It retries with backoff, and if it still fails the run parks in an exception queue with its state intact. It does not silently give up or start again from the top.

Can it act without a human?

Only where you decide it should. Approval gates are explicit, and anything touching money, external communication or deletion normally sits behind one.

How do we audit a decision later?

Every run is replayable. You get the inputs, the tool calls, the model responses and the output, stored and exportable, so a decision from three months ago can be reconstructed.

Who owns the workflows at the end?

You do. Repos, credentials and infrastructure transfer at handover. The workflows run on your accounts, not ours.

From the blog

Field notes

AI agent security checklist

The controls we run before an agent touches a real system.

Read it

Giving an agent a wallet

Spend limits, allow-lists and how agents get drained.