Skip to main content
Home is the first page you see when you open a project. It answers three questions in order: what is the agent doing, what does it need from me, and what has it done lately.
Every number on Home is read from your real data. If Zespan could not read something, the line says “I couldn’t read …” instead of showing zero.

Before your first trace

A new project shows one line: Send your first trace and I’ll start checking. with a link to Set up instrumentation. Home fills in as soon as data arrives. While you wait on the setup page, see First run.

First run

On the Connect your app step of setup, when the onboarding agent is on for your project, Zespan introduces itself with one line: I’m Zespan. I read every run and tell you what broke and why. Under it, the page shows what it is waiting for:
  • Waiting for your first run, with a pulsing dot and a running clock (Listening for your first run… 0:42). The dot is the only thing on the page that moves, and it stays still if your system asks for reduced motion.
  • An empty area, Your first run will draw here, one step per span, as it arrives. When a run arrives, the head reads Your first run arrived and the clock and this line go away.
Everything else on the step is where it always was: the key panel (paste or rotate), the sample trace, Continue once a run arrived, Skip and finish setup, and I’ll connect later. Check now appears only while Zespan cannot reach the API to look for your run. Screen readers hear a short status line only when the listener’s state changes, never on a clock tick: No run yet., Events are arriving., Your first run arrived., or, when nothing has come in after the listener’s wait (two minutes), Nothing after 2 minutes. Here is what to check. Check now keeps keyboard focus while it reads, and a line inside the head says whether the API answered. When you send the sample trace, a visible line says whether it was sent and takes focus.

What Zespan learned

On the Finish setup step, with the onboarding agent on, a card above the thresholds shows what Zespan read from your first runs. It lists only names, counts and step states, never the content of a run:
  • Agents, with an arrow for who delegates to whom (intake_agent → payout_agent).
  • Tools and how many calls each had. A count Zespan could not read says calls not available, never 0.
  • Instrumented as, the SDK and version your runs name.
  • Your first run, drawn as one block per step, linking to the run.
  • Questions · your answers pick the detector packs. Zespan asks only what it cannot know, for example whether a tool moves money. Yes, it moves money (or Yes) and No save your answer; Not sure saves nothing and the question stays open. A question about key fields lists candidate fields to tick, then Save. A question about an unclear cause offers Open a labelling queue. The line under the question repeats what the server answered: a detector pack is named as turned on or in shadow mode only when it really changed, and otherwise the line says no pack changed. Only an owner or admin can answer, and anyone else sees the question read-only with where to answer it, on Detector packs.
  • What I will watch and What I can’t see yet, with Open Instrumentation for the gaps.
  • First verdict, with Go to Home. If your first run had a failed step, it names the operation and says one run is not an issue yet. If no step failed, it says so without calling the run healthy. Before any run has arrived it reads No verdict yet with how long this step has been waiting. If a run has arrived but has no verdict yet, it says Your first run arrived; I’m still reading it. If Zespan could not read all of the run, it says so, and never shows the waiting clock. The wait stops counting after 30 minutes, when Zespan stops checking (Stopped listening after 30 minutes.), or while the last check failed, and the line then says so instead of claiming it is still listening.
While the card loads it shows a labelled placeholder, never No verdict yet. If Zespan could not read part of it, the card says some parts could not be read and shows not available for those parts. Nothing flagged so far under What I can’t see yet appears only when the instrumentation read succeeded and found no gap. If a later refresh fails, the card keeps what it read and offers Retry. The thresholds form, the role question and the next-step cards below it are unchanged.
Without the onboarding agent, the step looks and works as before. The first-run head appears only for projects where it is switched on.

The new Home

In the new interface, Home is one column and reads like a note from a colleague. It answers four questions in order: what am I doing, what do I need from you, what did I do for you, and what is coming. Zespan writes it in the first person (“I checked …”) and only says what it really read.
The new Home is part of the new interface. Until your project uses it, Home keeps the earlier layouts described further down this page. Nothing was removed: the loop summary moved to the sidebar’s Maintenance loop group, and the logbook moved into Done for you.

The status pill

The pill sits in the top bar on every page and says in one word what Zespan is doing. Hover it for the full sentence. Times are shown in your own time zone, with UTC on hover.
Silent is about your agents. Paused is only about Zespan’s own investigations hitting a limit.

Headline

Under the status you see one sentence that says the most important thing right now, and at most one button. Zespan picks it in a fixed order, so the same data always gives the same sentence: If Zespan could not read something it needs to be sure, the headline says “I could not read everything just now, so I won’t say all is well.” with Try again. It never calls the project quiet on a guess, and never says “All good” when there are no runs to judge. When the limit sentence is not the headline, for example because something critical ranks above it, it moves to the line under the headline. The line under the headline can also say when the next check happens, and it only does so when that check is really scheduled.

My day

Four rows sit directly under the headline. They are always there once the project is set up, and an empty row says “nothing”. If the whole block could not be read, it says “I could not read my day just now.”

Needs you

What only a person can do, grouped by kind. Each group has a first-person title such as “I need an owner to approve 2 tool calls.” or “I drafted 1 guardrail for you to decide.” Up to three groups are shown. The first group is open, with its main button, so you can decide without leaving Home. It works exactly as it does on Needs you, including the confirm steps. Batches for decisions come first. A batch that asks for your labels (“I need your labels on 2 issues.”) is always last, and a batch of code fixes reads “I drafted 1 fix for you to review.” A code fix card has a Review the fix link to the issue’s fix path, not an approve button. Before I ask for labels I try to investigate the cause myself, within your run limits; see I try first, then ask. In Running, such a run says “because I could not tell the cause yet.” A model test is one batch too: “I have 1 model test for you.” Before it runs, the card says what I would test and, for owners and admins, the most it can cost (“for up to $0.02”); Run the test opens a fresh estimate to confirm. When the cost cannot be priced, the card gives the call count and no dollar figure. After it finishes, the card gives the result in checks passed (“gpt-4o-new passed 8 of 8 checks; gpt-4o-old passed 6 of 8.”) and, where Zespan manages the model, a Switch triage to gpt-4o-new button that asks first. See Test the successor. A group can carry a line on what it is worth, for example “It would have stopped 9 of 14 failing runs in the replay.” Zespan writes that line only when it has the numbers, and leaves it out otherwise. If the headline already names an item, the group marks it “above” instead of repeating it. The section is hidden when nothing waits, and says “I could not read this list just now.” when it could not read the queue.

Critical now

What is breaking right now: an incident, a critical alert, a failed issue ranked by what the tool does, a failing agent, a model that has already been retired, an agent that did not answer my canary or failed its checks twice, or a tool whose answer changed shape and broke, refused calls or ran out of credits (Tool drift), or one of your agents that I stopped, that is failing fast, that dropped below its target or that is past its daily budget (Agents). An agent I stopped ranks first and takes the headline: “I see this now: I stopped refund_agent at 14:05 because it used its whole daily budget. It can run again at 00:00, or you can lift the stop.” Owners and admins get Lift the stop; everyone else gets See the agent and is told an owner or admin can lift it. No dollar figure appears in the headline or here. A fast burn reads “I see this now: refund_agent is failing 45 of 120 calls in the last hour, against its 99% target. At this pace its 7-day error budget is gone in 3 hours.” A breach, a slow p95 and a budget overrun with no stop in force read in the past tense with the time they began. When one of these says “past its daily budget” because I cannot stop that agent, it says why (an older SDK, OpenTelemetry only, an agent name not set in code, or an SDK in log mode) and does not claim the stop is off. When my agent checks are not running, these rows are worded in the past with the time I last read them, never as “now”. This section is hidden unless something is critical, and there is no “nothing critical” box. If one of the checks behind it could not be read, it says so in one line instead of staying quiet. A failed issue row also says how many customers it reached, for example “Failed 12 times, 4 customers, last seen 45m ago”, when your app sends user ids; otherwise the customer count is left out.

Coming up

Dated risks, soonest first, up to five. These are an error budget about to run out (“At the last 6 hours’ pace, refund_agent uses up its 7-day error budget on Nov 24.”, shown only with at least 200 calls and never for an agent already below its target), model retirements (the date, the model, how many agents and prompts still call it, the successor when one is named, and “I tested the successor on Nov 21.” once a model test has finished), a key expiry date you gave me or I read from a key (“The Stripe key key, used by refund_agent, expires on Dec 1.”), and a quota with its projected run-out date. A date you declared is shown as the day you typed, with no clock and no time zone shift. The section is hidden when nothing is coming. When a key or quota already has a card in Needs you, it is not repeated here. If I could not read the model lifecycle I say “I could not read what is coming up.” Open Model lifecycle for the full list.

Agents

One line for each of your agents, worst first, up to six, with All agents linking to the agent list. It sits after Coming up and is not drawn until agent targets are on and I have read them. An agent with no target says “no target set”. I never make one up. A version I derived reads “version z-ab12cd34 (derived by me)”. If I tied a drop to a version and its card is not shown in Needs you, the row says so once and links to the versions. If I could not read your agents it says “I could not read your agents just now.” It never reads as “all met”.

Done for you

What Zespan did, newest first, up to eight rows, with All entries linking to Activity. Rows since your last visit come first and carry a New label. When nothing is new it says “Nothing new since your last visit.” and lists the latest entries under Earlier.
  • Each row is a plain sentence, with the cause only when Zespan has one: “I grouped 14 runs into one issue because they share a failing tool call.”
  • Routine work is rolled up into one line per kind per day, for example “I checked 6 changes from today.” A roll-up has no link of its own. Use All entries to see the rows behind it. When Zespan read only part of the log, the count says “at least”.
  • Whether a fix you applied is holding shows here too, with its outcome in words.
  • A stop and a miss are rows of their own, never rolled up: “I lifted the stop on refund_agent: its new day started.”, “Lifted the stop on refund_agent.” when a person did it, and “refund_agent is back within its 99% target.” A stop or a miss that Critical, Coming up or Needs you already shows is not repeated here.
  • My canary shows here too: a roll-up such as “My canary ran 6 times on Nov 21 (UTC) on refund_agent; all passed.”, a first failed check with when I will run it again, a passing run whose answers changed shape, and a canary I stopped at its daily limit. A day with any failure has no roll-up, so a failure is never folded away.
  • A finished model test and a test that stopped early are rows of their own, never rolled up: “I tested gpt-4o-new against gpt-4o-old on 20 calls: 37 of 40 checks passed on gpt-4o-new.” While the result card still waits in Needs you, the row is not repeated. A fix I drafted on my own is a row too.
  • Once a day Zespan writes one “rounds” entry that says how many times its checks ran, how many new runs and issues it saw, and how many of your models it compared against its retirement list.
If it could not read its log, it says “I could not read my log just now.”

How I checked

A collapsed section whose summary line is always visible, for example “How I checked · 1,240 runs · 2 things not checked · 3 of 5 investigations used”. Open it to see:
  • What was checked: “I checked 1,240 runs and 38 scores today.”
  • Not checked: what Zespan could not look at, such as tool arguments that were not captured, answer quality when there is no LLM provider, or something it could not read.
  • My canary: “My canary checks 2 agents; the last run was at 14:12.”, the agents with no canary, and a note when your agent’s own spans do not carry my mark.
  • My agent checks: “I checked 4 agents against their targets at 14:12.”, the agents with no target (“Targets for intake_agent (no target set).”), the agents whose latency I did not measure, a read that ran out of time, and a note when my agent checks are not running. If I could not read them it says “I could not read your agents’ targets just now.”
  • My fix drafting: “I look for proven issues I can draft a fix for every 15 minutes. Today I drafted 3 of my 5 allowed.” When I could not read the count it says “not read”, never 0. Absent when auto-drafting is off.
  • How often: “My checks run every 10 minutes; the last one was at 14:12; the next is at 14:22.” If the checks are not running it says that instead.
  • Zespan’s own work today: “Today I used 3 of 5 automatic investigations and judged 120 of 200 runs. Limits reset at 05:30.” If investigations are off for the project, it says “Investigations are off.”
Owners and admins also see Zespan’s own spend in dollars, as “That is $1.20 of the $3.00 daily cap you set.” Everyone else sees counts only. Dollars never appear in the headline or the pill.

Ask Zespan

A box at the bottom, “Ask Zespan about your agents”. It opens ZespanPilot with your question filled in. Nothing is sent until you press send there.

Daily limits for automatic investigations

Zespan limits its own automatic investigations so it cannot run up cost unnoticed. The project has a daily limit on automatic runs, and the organization and Zespan itself have limits too. When a limit is reached, new automatic investigations wait, and Home says so in the pill, the headline, Running and How I checked. Limits reset at 00:00 UTC, shown in your own time zone. After the reset, Zespan starts the waiting issues that still qualify. An issue that has since gone stale, or has been closed, is not started. Owners and admins can change the project limit from the headline’s Change limit button. See Agent runs for the limits.

How Zespan keeps Home honest

Zespan prefers saying “not read” to guessing. Any number that could not be read is shown as “not read” and never as 0, and Zespan does not call a project quiet or clear while something it depends on is unread.
  • Next times are real. A “next check” or “next” time appears only for a job that is scheduled and running. Otherwise no time is shown.
  • Checks are not findings. A last-check time proves Zespan’s checks ran. It does not claim the checks found nothing for your project.
  • Silence is not Zespan’s silence. When your agent sends nothing, Home says so and shows that Zespan is still on watch. When your SDK key has been seen, or refused, recently, it says which, and the reason when it refused a request.
  • Each fact once. Something the headline already said is not repeated in a section below it.
  • No pulsing without proof. The only moving dot is a step of an investigation that is running right now.

The Today digest, queue first

On the earlier layout of the new interface, Home reads top to bottom in this order, and the first decision is on screen without scrolling.
1

The Today digest

A short note written by the agent. The lead sentence says what I checked and how many things need you. Under it, one line on today’s runs (or that none have arrived, and when the last one did), one line on what is waiting on you, and a line on what I did not check. If I could not read today’s runs, that line is left out instead of showing zero. If I could not read the capture coverage, the “not checked” line is left out too, and I never say nothing is missing.When runs are available for the project, two more lines can follow. Last run names the most recent run, such as “Zespan · run done Sat 08:35”, and links to it. It is left out when there has been no run. When Failed issues are waiting for the run limit, a line says how many, for example “2 Failed issues are waiting for tomorrow’s run limit”. It is left out when none wait, and also when I could not read the count: I do not show zero.
2

The first item in Needs you

The top item of Needs you is a full card with its one main button, so you can decide or apply without opening another page. It works exactly as it does on Needs you, including the confirm steps for applying a fix and publishing a postmortem. The count beside the heading is the same number as the chip in the top bar.
3

The loop, folded

Only the stages that have something waiting are listed, each with how many items wait. The other stages sit in a fold labelled “Other stages, nothing waiting”. If I could not read what waits in each stage, Home says so and does not claim nothing waits.
4

The logbook preview

The eight latest entries, with the time, the stage and a link when the entry has a page. The rows are not numbered in this preview. Open Activity for every entry.
5

Ask

The same box that opens ZespanPilot with your question filled in.

A fresh project

When the project is set up, has no open issue and nothing waits on you, the queue area shows what I checked and what I did not: today’s run count, the capture gaps I could not check, and a note that findings land in Needs you as a Decide, an Answer or an FYI. The same note appears at the top of Needs you when its queue is empty, and at the top of Activity. If a number could not be read, the note says so.

The maintenance loop strip (previous layout)

A row of cards, one per stage of the loop, each with a short summary and a link to that stage: Click a card to open the stage.

Needs you and Recent activity

Below the strip, Home shows the top three items from Needs you, with a link to see them all, and the latest items from Activity. You can approve, reject, or act on an item right from Home. At the bottom, an Ask me about your agents box opens ZespanPilot with your question filled in. Nothing is sent until you press send there.

What the agent tells you

Most Maintenance pages start with a note headed Zespan agent. It is the agent’s lead on that page.
1

Read up to three points

Each point says what the agent found, in plain words, with the real numbers from this page’s data. Nothing is invented.
2

Take the next step

Each point comes with an action: a button that opens the right page, or one that records your decision, such as marking a cost recommendation as applied.
3

Or enjoy the quiet

When I checked everything on the page and nothing needs you, the note shrinks to one line: “Nothing needs you here. I checked 120 runs today.” If there were no runs today it tells you when the last one was. If I could not check everything, I say that instead and never claim it is quiet. Pages with nothing for me to check show only the run count.
The note appears on the Observe, Detect, Fix, Prevent, Evaluate, Optimize, and Govern tabs. It does not appear on Home, Needs you, Activity, Ask the agent, the other project Settings pages, or on a single trace or issue.
Times in the note use your own time zone, so “today” means your today.

Next steps

  • Needs you: clear the decisions waiting on you
  • Activity: see what the agent did
  • Agent runs: watch an investigation, and what the run limits are
  • Model lifecycle: see every model that is retiring, and test its successor on your own calls
  • Guardrails: the fixes I draft on my own
  • Canary: scheduled test runs of an agent, and what Home says about them
  • Tool drift: a changed tool, a key about to expire, a quota running out
  • Agents: targets, the daily stop and the version record behind the Agents row
  • Notifications: get the daily digest by email or chat
  • Navigation: how the sidebar is grouped
  • ZespanPilot: ask the agent a question