Agents · the long version

Autonomous,
on a leash.

Set an agent loose to plan, call tools, and do the work — with hard limits on how many steps it takes, how long it runs, and when to stop, all enforced by Catalyst rather than left to the model's good intentions. Every step is visible. You can pause, step in, or stop it instantly at any point.

Safety, by default

Caps the model can't argue with.

The agent keeps going until the model stops reaching for tools — bounded by hard limits Catalyst enforces, not by the model's self‑restraint. Three dials, every run.

max iterations

Default 5 · hard ceiling 50.

How many rounds of planning and tool‑calling the agent can run. Most real tasks finish in three to seven; the hard ceiling of 50 means even a worst case can't run away.

time budget

300 seconds, configurable.

Total real‑world time the run is allowed. A slow tool call still counts against it. When time's up, the agent stops where it is and you get whatever it finished so far.

stop when

Optional finish phrase.

Give it a word or phrase to watch for; the moment the model's answer contains it, the run ends — even if it had rounds to spare. Handy for "stop as soon as you have the answer."

transparent execution

Every step. Every tool. Every error.

Agents are usually black boxes — input goes in, output comes out, the middle is a guess. Not here. Catalyst shows every step as its own card: which tool the model reached for, what it sent, what came back, how long it took, and whether it failed. Click any card to open the full details.

  • Per‑step cards. Step number, tool name, inputs, result, status, time taken. No endless wall of text to scroll through.
  • Errors recover, not crash. Ask for a tool that doesn't exist? The agent gets a clear "tool not found" back and tries again — instead of the whole run dying.
  • Prompt caching kicks in on round two. The instructions are identical every round, so the model reuses the cached copy and you pay only for what's new.
  • Stop mid‑run. The stop button reaches all the way up to the model — nothing keeps running in the background, and the meter stops with it.
agent run · portfolio brief · iter 3/5
1
tool · brave_search
{ query: "NVDA earnings Q3" }
5 results · 412ms · cache: miss
2
tool · fetch_quotes
{ tickers: ["NVDA"] }
1 quote · 88ms · cache: read 7.2k
3
tool · python.exec
{ code: "compute MA + delta" }
running…
4
awaiting prior iteration
3 / 5 iterations · 76s / 300s · $0.04 spent
What people build

Three patterns. Many shapes.

Agents come with three starting patterns proven in production. Start from one and mix in your connected tools as you go.

research

Search agent.

Take a question, search the web and your own docs, read the sources, and pull it together. Capped so it doesn't dig forever.

analysis

Python analyst.

Take a dataset — a spreadsheet, a file, a query result — run Python in a private sandbox with pandas / matplotlib, and return a chart plus a written summary. It keeps going until the answer holds up to its own checks.

briefing

Portfolio brief.

Pull live market data, crunch the numbers in Python, format the result, and email it out. The go‑to setup for "every weekday morning" scheduled agents.

Bounded autonomy

Trust the model.
Verify the budget.

Agents earn trust through transparency, not promises. Caps the model can't bypass. Transcripts you can audit. A stop button that actually stops.