How the agent works

Excelente runs an agent loop, not a single request. Understanding the loop explains most of what you see in the transcript, including why it pauses where it does.

The loop#

Every message you send starts a turn sequence:

1. Excelente sends your message, the conversation so far, and the list of available tools to your model.
2. The model streams back text, tool calls, or both.
3. Each tool call is checked against its permission level. Read tools run immediately. Write tools stop for your approval.
4. The tool result goes back into the conversation.
5. Repeat from step 2 until the model stops, or until the turn limit is reached.

A "turn" here is one round trip to the model, so a build that writes eight ranges and takes four screenshots has burned at least a dozen turns before it says a word to you.

What the agent brings to each turn#

Alongside your message, the system prompt carries:

• The workbook outline, which sheets exist and roughly what is in them.
• A.CRE house modeling conventions, injected in full on every turn so formatting and formula discipline stay consistent without you asking.
• Any skills you have enabled, as short summaries the agent can expand on demand.
• Instructions from any connected MCP servers, plus their cached data catalogs.
• Workbook memory, if this workbook has any.

Everything else is pulled in only when the agent asks for it. See What the agent can see.

How long it works before checking in#

Excelente pauses after a stretch of work rather than running unbounded. You set the stretch under Settings → Advanced → How long the agent works:

When it pauses you get a notice and a Continue Working button:

Excelente paused after 200 steps, its limit for one message. Nothing it has done is lost.

Clicking it picks up where it left off. You can also type what you want instead, which redirects rather than resumes.

The same prompt appears when the agent stops with items still open on its checklist, worded to say how many.

The included model is pinned to 200 turns and does not show this setting.

Steering a run#

The composer stays live while the agent works. Anything you type and send is queued and delivered at the agent's next tool boundary, so you can redirect mid-build without stopping it:

⏳ Use 4% growth instead — queued, delivers at the next step

Queued messages show a ✕ to withdraw them. If a run ends while something is still queued, the text goes back into the composer rather than disappearing.

To stop outright, click the Stop button, which replaces Send while a run is in flight.

Compaction on long conversations#

Conversations have a working budget of roughly 200,000 tokens. Past that, Excelente folds the older turns into a structured summary and carries on, which keeps a long session coherent instead of truncating from the top.

Three tool results are never compacted away, because losing them mid-build causes the agent to repeat expensive work or forget your workbook's rules: load_skill, read_skill_resource, and read_workbook_memory.

Compaction runs once per long session. If you have a key, you can point it at a cheaper model under Settings → Advanced → Summary model, since summarizing does not need frontier reasoning.

Where verification fits#

In Work mode the loop does not end when the last cell is written. The agent screenshots what it built, reads the screenshot, then spawns a Reviewer sub-agent for an independent check. Only then does it summarize. See Sub-agents and verification.

Reading the transcript#