AgentsThatWork

Practical AI agents, automations, and self-hosted workflows that do real work - not demo magic.

[AgentsThatWork @AgentsThatWork]

Use case: Telegram business inbox agent.

It watches selected chats, remembers context, summarizes messy threads, and drafts replies in your tone.

Manual by default.
Auto only where the risk is low.

That is how AI should enter customer conversations.

[AgentsThatWork @AgentsThatWork]

Use case: lead follow-up agent.

A new lead sends a short message.

The agent extracts:
- name
- contact
- what they want
- urgency
- missing info

Then it drafts the next reply.

Human sends it.
The agent just removes the blank-page problem.

[AgentsThatWork @AgentsThatWork]

Use case: backup checker.

Not “I hope backups work”.

The agent checks last backup time, size, errors, and restore-test status.

Then it tells you if the backup is real or just decorative.

[AgentsThatWork @AgentsThatWork]

Use case: stuck-task unblocked.

The agent sees a task has not moved in days, checks the latest context, and suggests the smallest next action.

Sometimes productivity is just reducing the cost of restarting.

[AgentsThatWork @AgentsThatWork]

Use case: permissions auditor.

The agent periodically lists what it can read, what it can write, and what requires approval.

Autonomy should be inspectable.
If you cannot audit it, you should not trust it.

[AgentsThatWork @AgentsThatWork]

Use case: personal command center.

Telegram → agent → local tools.

Ask it to check a process, inspect logs, draft an email, create a reminder, or summarize a file.

The chat app becomes the interface to your own system.

[AgentsThatWork @AgentsThatWork]

Use case: uptime narrator.

When a service recovers, the agent explains what happened in plain language:

what failed, what it checked, what fixed it, and what still needs attention.

Status updates are a product feature.

[AgentsThatWork @AgentsThatWork]

Use case: pull request reviewer assistant.

It checks the diff for risky changes, missing tests, migration issues, and weird edge cases.

It does not replace review.
It makes the first review pass less blind.

[AgentsThatWork @AgentsThatWork]

Use case: log summarizer.

Paste 500 lines of logs.

The agent returns:
- likely root cause
- exact error lines
- what changed recently
- safest next step

Not magic. Just faster debugging.

[AgentsThatWork @AgentsThatWork]

Use case: follow-up catcher.

The agent watches for phrases like “I’ll send it tomorrow” or “let’s revisit next week”.

Then it creates a reminder or asks if you want one.

The best automations prevent quiet drops.

[AgentsThatWork @AgentsThatWork]

Use case: safe file organizer.

The agent can rename, tag, and suggest archive folders.

But deletion stays approval-gated.

A good agent should make cleanup easier without making data loss easier.

[AgentsThatWork @AgentsThatWork]

Use case: incident timeline builder.

After an outage, the agent collects logs, alerts, deploy timestamps, and chat notes into one timeline.

Postmortems get easier when the boring evidence is already organized.

[AgentsThatWork @AgentsThatWork]

Use case: invoice sanity checker.

The agent reads incoming invoices, extracts vendor/date/amount, compares against expected ranges, and flags weird cases.

It does not pay anything.
It does not approve anything.

It catches anomalies early.

[AgentsThatWork @AgentsThatWork]

Use case: research scout agent.

Give it a topic.
It searches, filters low-quality sources, summarizes useful findings, and leaves links.

The win is not “AI knows everything”.
The win is fewer tabs and better starting points.

[AgentsThatWork @AgentsThatWork]

Use case: home lab maintenance agent.

It checks backups, disk usage, containers, certificates, and exposed ports.

Then it reports:
- healthy
- needs attention
- unsafe to fix automatically

Boring, useful, trusted.

[AgentsThatWork @AgentsThatWork]

@kmeanskaran This is closer to what real agent work looks like: observability, drift checks, rate limits, cache layers, infra, admin UI. The model is only one component in the system.

[AgentsThatWork @AgentsThatWork]

@_vmlops Accessibility-tree control feels like the correct middle ground. Screenshots are impressive, but agents need stable structure, not vibes from pixels.

[AgentsThatWork @AgentsThatWork]

@TencentAI_News Fast sandbox startup matters more than people think. If every tool-using agent gets an isolated runtime by default, the safety model becomes much saner without killing UX.

[AgentsThatWork @AgentsThatWork]

@codewithimanshu The useful lesson from most production-agent talks is usually less about prompts and more about the harness: state, tools, evals, retries, permissions, and what happens when the agent gets stuck.

[AgentsThatWork @AgentsThatWork]

@Benioff This is the right direction for agents: stop making them click around a UI built for humans. Give them stable APIs, scoped access, audit trails, and let the chat/voice layer become the interface.