AI agent observability, every step on the record.
Every Arclyx run is recorded step by step: what the agent decided, each tool call and what came back, the tokens it used and what it cost. Each agent's page charts its run history, and agents that need attention are flagged on your home page, so you see a problem before someone tells you about it.
- Gmail · searchis:unread newer_than:1d
- Gmail · read14 messages
- ThinkingSorting into Needs you, Can wait, FYI
- Email · send to meToday's inbox: 5 need you
Why do AI agents need observability?
Because a wrong answer reads exactly like a right one. The only way to know an agent did the job properly is to see how it got there: what it searched, what it read and what it skipped. And when it runs on its own, the record is the only witness.
Without it
- You see the answer, not how it got there
- A slow or costly agent hides in the total
- You hear about a broken run from a colleague
With Arclyx
- Every step and tool call, in order
- Tokens and cost for each run
- A Needs you flag the moment one stops short
What does Arclyx record for every agent run?
The whole run, live while it happens and kept afterwards.
Every step
Each round: what the agent decided and what it did next.
Every tool call
The tool, what it asked for and what came back.
Tokens
Input, cached, output and reasoning, for the run and for each model call.
Cost
The run's price and a line for every call, on paid plans.
Helpers
When an agent hands work to helpers, their runs are recorded too.
What does a recorded run look like in Arclyx?
Every run has its own page: what happened in a sentence, the numbers, a round-by-round replay and the model calls behind it. The agent's page charts all its runs together.
ActivityInbox triageRun
@inbox-triage · by schedule · today 8:00 AM
It sorted 14 emails and flagged five that need you today.
Done after 38s
- Took
- 38s
- Rounds
- 5
- Tool calls
- 7
- Tokens
- 41K
- Price
- $0.02
Round 3 of 5
Sort into Needs you, Can wait and FYI
Thinking
Billing
Model calls across this run
| Round | In | Cached | Out |
|---|---|---|---|
| 1 | 8.2K | 0 | 96 |
| 2 | 11K | 7.9K | 140 |
| 3 | 12K | 10K | 310 |
| 4 | 4.6K | 3.9K | 210 |
| 5 | 2.1K | 1.8K | 160 |
AgentsInbox triage
Every run
Pick a dot to open it
Done Needs you
What should you trace in an AI agent?
Everything between the request and the answer. A run can finish without an error and still have called the wrong tool or read the wrong data. These are the parts of a trace that catch it, and Arclyx records each one.
Each decision
What the agent chose to do in each round, and what it did next.
Each tool call, with its input
The tool, what it asked for and what came back, so a wrong search shows up.
What it read
Documents and pages it searched or opened, as tool calls you can open.
Handoffs to helpers
The brief and model each helper was given, and the helper's own run.
Tokens and cost per call
So one expensive step stands out from the rest.
How do you spot a problem across many runs?
Each agent's page charts its runs as dots, by duration, tokens or price. A run that took twice as long or cost twice as much stands out; click it to open the run. Agents whose last run stopped short show as Needs you on your home page.
- Every run as a dot, by duration, tokens or price
- Flagged on Home when a run stops short or a connection breaks
- Every run across all your agents, in one place
- Every schedule and webhook, with pause and delete
How do you debug an AI agent run that went wrong?
Start from the run, not the answer. Find the round where it went off track, see what the tool gave back, and fix the cause: the instructions, the tools, or a broken connection. Then run it again and compare.
- A Needs you flag on Home, or a dot that stands out on the run-history chart
- Every round, tool call and result, in order
- The round where it searched for the wrong thing, got nothing back, or kept going
- Change the agent in plain words with Improve, or reconnect the tool
- Run it now, and compare the new run with the old one
How do you see what your AI agents cost?
Settings → Usage shows the last 30 days of spend or calls with a daily chart, and how close you are to each limit. Admins can switch to everyone's usage. Every run's own page shows what that run cost.
- Spend or calls, with a daily chart
- Meters for each limit you've set
- Admins see usage across the workspace
- The price of each run, call by call
Questions about AI agent observability.
What is AI agent observability?
Seeing what an agent did on every run, not just what it answered: each decision, tool call, result, token and cost, so you can tell whether it did the job properly.
Is observability the same as monitoring?
Monitoring tells you something went wrong; observability lets you see why. Arclyx flags problems and keeps the full record of every run. Our post on the observability gap goes deeper.
Do I need a separate AI agent observability tool?
Not for agents you run in Arclyx. Observability is built in: every run is recorded step by step with nothing to install or set up.
How is observability different from evaluation?
Observability shows what an agent did on real runs. Evaluation scores it against a set of cases. In Arclyx you check a draft with test runs, then watch live runs in their history.
Can I see what each run cost?
Yes, on paid plans each run shows its price and a line for every model call. On Free, runs show tokens and usage shows your share of the allowance.
Where to go next.
See every step your first agent takes.
One sentence in, a working agent out. Join the beta and we'll set you up.