Agent configuration and operating experience

My personal AI assistant

I configured existing agent software for messages, notes, web tools and scheduled tasks. The story covers operational failures as well as the configured capabilities.

Setup source review: 11 September 2026. This description does not certify current operation.

What was built?

The work was the agent’s operating environment and behaviour, not a new language model or a runtime built from scratch. The setup connected one permitted user’s messaging channel, a local container, persistent memory files, and browser, terminal and file tools. Recurring activities became scheduled jobs. Persona, instructions and skills shaped how tools were used.

How the work happens

A message in. Prepared work out.

6 steps
WorkflowHuman decisionWhen something needs attention
  1. 01Your task

    “Summarise my day.”

  2. 02Set the scope

    Which day and which sources are permitted?

  3. 03Read permitted context

    Gather the relevant notes and agreements.

  4. 04Collect the useful facts

    Separate completed work from open questions.

  5. 05Prepare the summary

    AI turns the retrieved facts into a readable overview.

  6. 06You choose what comes next

    A summary does not send mail or edit the calendar.

    Human decision
An illustrative Hermes task. A summary alone needs no permission to send or change things.

Workflow and review boundary

  1. Message

    The user gives a task through the allowed messaging channel.

  2. Context

    The agent uses instructions and local memory files. Persistent files do not guarantee that every earlier detail will be retrieved when needed.

  3. Tool use

    A task may require reading a page, changing a file or executing a command. Tool availability and successful task completion are different things.

  4. Review

    External actions require a distinction between preparation and approval. The boundary shown here is an operating requirement, not proof of a technical safeguard that blocks every action.

  5. Result

    An answer or file becomes an inspectable result. For scheduled work, inspect the latest run rather than merely checking that the job is enabled.

What operation revealed

Provider balance is part of reliability.

One scheduled task stopped when the external model account ran out of funds. The local server remained online, but the task did not finish. A resolution had not been retested in the source review.

A symlink does not resolve an access boundary.

Two scripts sat outside the scheduler’s permitted directory. A symlink workaround failed because the guard resolved the actual path. The jobs remained broken at review time.

Configured does not mean connected.

Some additional channels and services appeared in configuration but were not operational. They cannot be described as delivered integrations.

My contribution

I defined the assistant’s role, tools and everyday use, and operated the installation. AI helped with instructions, skills and scripts. The underlying runtime is third-party open-source software.

Data boundaries

Files and tools were local, but conversation content and context went to an external model service. Telegram is also external. Local hosting does not mean entirely local processing.

What the review checked

3 configured weekly jobs

Source review of one personal-agent installation, 11 September 2026. All three last inspected runs had failed. This measures configured scope, not reliability; there was no automated test suite.