Reflex · Deployed agents

Agents that work without being asked

Describe a job in a sentence and Reflex turns it into an agent that runs on a schedule, across your systems. Some jobs are the same steps every time. Others need judgement. Reflex has a kind of agent for each.

01

Some work should not wait for a question

A chat is good when you know what to ask. A lot of useful work is the question you would ask every morning if you had the time.

Are all our ERP users set up in the warehouse system? Did any new orders arrive overnight? Which accounts have not been used in months? Is the warehouse system still secure? Each of these is easy to answer once. The value is in answering it every day, without anyone remembering to.

A deployed agent does exactly that. It is given a job and a schedule. It wakes up, reads what it needs from your systems, decides what matters, acts on it, and records what it did. Then it does the same again tomorrow.

1Wake upThe agent starts itself, on its schedule, day or night.
2ReadIt calls the tools it needs, on the systems it was given.
Standing goal“Keep late orders and lost revenue as low as possible.”
3DecideIt works out what matters: by fixed steps, or by reasoning.
4Act and recordIt emails, reports or updates, and the run is kept in the history.

then the loop runs again

Figure 1. Every deployed agent follows the same loop. The two kinds differ at step 3.

02

Made by asking

You do not draw a diagram or write a script. You write what you want, the way you would ask a colleague, and Reflex works through four stages with you.

First it understands: it looks at your connections, finds the tools that fit, and asks about anything the sentence left open, such as what to do when a step fails. Then it writes the requirements back to you in plain language, including the exact text of any email it will send, and waits for you to approve them. Only then does it build. When it is ready, you can run it at once or put it on a schedule.

Reflex designs the agent from the descriptions of your tools. It does not need to see your data to do it.

03

Operational agents: the same steps, every time

When the job is a procedure, Reflex builds an operational agent. What it produces is a plan: a short chain of steps, each one a tool call, a query, a condition or a message.

In the example, the plan reads users from the ERP and from the warehouse system at the same time, finds the ERP users with no match, counts them, and sends an email only if the count is above zero. The comparing and counting are done with queries, so the result is exact. The step that sends something out is marked as one that writes.

The plan is checked before it is saved, and you can read every step of it. Each run follows it in the same order, so the answer today is worked out the same way as the answer last week. Read about Operational Mode.

04

Reasoning agents: a goal, thought through each run

Some jobs cannot be written as fixed steps, because the next step depends on what the last one found. Run a security assessment and, for each critical finding, check whether it has already been reported. How many findings there will be, and what to search for, is not known until the scan has run.

For a job like this, Reflex builds a reasoning agent. It writes a prompt, not a plan, and hands it to an agent you choose. You test it exactly as it would run on a schedule, and read what each agent did: which tools it ran, and what it handed to the next. When the result is right, you save it.

Each run, the agents plan again from the prompt, so they can follow whatever they find. Read about Reasoning Mode.

05

Which one, when

The rule of thumb is simple. If you could write the steps on a card and hand it to someone, it is an operational agent. If you would say “look into it and tell me what you find”, it is a reasoning agent. Most teams end up with many of the first and a few of the second.

Compared onOperational agentFixed stepsReasoning agentJudgement each run
What you describeA job with known stepsA goal that needs judgement
What Reflex buildsA plan of steps, checked and savedA prompt for an agent, tested
What happens each runThe same steps, in the same orderThe agents plan again from the prompt
How it decidesQueries and conditionsThe model reads results and reasons
Best forChecks, comparisons, counts, regular reportsInvestigations that depend on what is found
ExampleWhich ERP users are missing in the warehouse?Scan for security issues, then check which are already reported

06

A use case: is the system really working?

Most monitoring answers one question: is the server responding? It pings the machine or calls a status URL, and if something answers, the light is green. Anyone who has run a warehouse or an ERP knows the light can be green while nobody can release an order.

A deployed agent can check the way a person would. On its schedule it makes real but harmless calls through the same tools everyone else uses: it signs in, reads a small piece of live data, times the answer, looks at whether the background jobs and interfaces are moving, and checks how fresh the latest transaction is. Nothing is changed in the system.

The result is a picture of health across several dimensions, not a single up or down. When one of them slips, the agent says which one, and tells the people who need to know.

Warehouse system healthEvery 5 minutes · last check 2 min ago
Degraded
  • Sign-inOK

    A test login is accepted
  • Live readOK

    Today's orders came back
  • Response timeSlow

    9.2 s, usually under 2 s
  • Background jobsStopped

    Allocation job has not run since 01:00
  • InterfacesOK

    Last ERP message 4 min ago
  • Fresh dataOK

    Newest transaction 3 min old

Emailed to the warehouse IT team: allocation job stopped, responses slow.

Figure 2. One check, six dimensions. Each tile is a real, harmless call through a tool, so the agent reports what is wrong, not only whether the server answers.

This is an operational agent: the checks are the same every run. The dimensions are yours to choose, because each one is a tool your team has written for that system.

07

On a schedule, as you

Any deployed agent can be run on demand or scheduled: every few minutes, hourly, daily, on weekdays, weekly, monthly, or to a pattern of your own, in the time zone you choose. Before you confirm, you see the next runs listed, so there is no doubt about when it will fire.

A scheduled run uses your connections and acts as you. A step that says “email me” writes to you, and if the run itself fails, you are told by email.

08

Every run on record

Work that happens while nobody is watching has to be easy to check afterwards. The history shows every agent on a timeline: when each run happened, how long it took and how it ended. Any run can be opened to see what it did.

You do not have to read the timeline to find trouble. An agent whose last run failed, or whose scheduled run did not happen, appears in a list of its own with the reason in plain words, and leaves the list once it runs well again.