What a Morse performance brief is (and what it is not)
A performance brief is a ranked answer to one question: which of your production agents deserves attention first? Here is what goes into it, what stays out, and how teams use it week to week.
By The Morse team
Once a team has more than a couple of agents in production, the hard question stops being "what happened in this run?" and becomes "which agent should we work on this week?"
Most tooling answers the first question well. You can open a trace, read every model call, and see every tool result. What you usually can't do is look at one page and know which agent is getting better, which one is just getting more expensive, and which one you don't have enough evidence to judge yet.
That page is what Morse calls a performance brief. Here is what goes into it, what we keep out, and how a team uses it.
The loop: measure, prioritize, verify
The brief exists to run one loop.
- Measure. What happened to each agent in production? Morse shows labeled outcomes, spend, and sampled quality signals, each with its source, time window, and coverage.
- Prioritize. Which agent deserves attention first? Morse turns those signals into a ranked list, most urgent first.
- Verify. Did the change help? Morse re-measures the same signals in the next comparable window, and links back to the traces you need to inspect what moved.
Everything else in the product — cost attribution, scorers, context and memory profiling, deploy markers, the trace viewer — exists to make one of those three steps more credible or less work.
What goes into a brief
A verdict in Morse is a ranked brief built from four kinds of evidence:
| Signal | The question it answers |
|---|---|
| Labeled production outcomes | Did the agent do the job you defined as success? |
| Spend | What did it cost, including cost per successful outcome? |
| Production-success trends | Is the success rate moving up or down against the previous window? |
| Sampled quality signals | What do the quality checks say on the runs that were sampled? |
The important word in that last row is separate. Outcomes and sampled quality answer different questions. An outcome label says the business result happened. A quality score says a scorer liked the output of a sampled run. Morse never treats one as the other, and a sampled quality signal shows its sampling and coverage instead of presenting itself as the truth about every run.
Each agent in the brief carries one of four plain verdicts: Needs attention, Worth watching, Healthy, or Not enough labeled outcomes. That last one matters as much as the other three.
What a brief is not
We get asked about this a lot, so it's worth being blunt.
It is not a fused cost × eval score. There is no single number that blends dollars and pass rates into an "agent score." The ranking is an attention order built from explicit rules, not a weighted blend. Dollars can help decide which of two problems to look at first; they never get mixed into a quality number.
It is not a root-cause claim. A brief tells you where to look and what changed. When the evidence only shows that two things moved together, the brief says so and stops there. If a number moved and no deploy lines up with it, the brief states the change and does not invent a cause.
It is not an autonomous fix. The brief suggests one next measurement, inspection, or change. You decide whether to make it.
When the data can't support a conclusion, Morse withholds the item and says what has to arrive before it can rank it. Cost per successful outcome stays empty rather than showing zero when no outcome can be confirmed. We think a visible "not enough evidence" is more useful than a confident guess.
Who it's for
The brief is built for hands-on founders and engineering leads who run a live agent fleet. It fits best when:
- several production agents compete for the same engineering time;
- a release, model change, or spend increase keeps raising the question "which agents actually got better?";
- you have, or can define, what success means for each workflow; and
- quality review is sampled or partial, and nobody wants to run a full eval program just to pick this week's priority.
It is a poor fit for pre-production demos with no real traffic or labeled outcomes yet, or for anyone wanting a universal agent score or guaranteed root cause.
How a team uses it, week to week
The brief arrives every week in your dashboard. A typical week looks like this:
- Read the top of the list. The first item is the one with the strongest case for attention. Read its outcome, spend, and quality evidence side by side.
- Open the evidence if you need it. Every ranked item links to the traces behind it. Most weeks you open a handful, not hundreds.
- Make one change. A prompt, a model, a routing rule, a tool definition, a spending limit. One change keeps the next window readable.
- Check the next window. The following brief re-measures the same signals. If you accepted a recommendation, it comes back so the change gets re-checked rather than quietly assumed to have worked.
Morse reports the movement it observed and the coverage behind it. It doesn't claim your change caused an improvement unless the evidence supports that.
The launch offer: a founder-reviewed brief
At launch, the brief is also a concierge offer. A founder reviews every launch-period brief before you get it, and it is delivered within two business days of evidence readiness.
"Evidence readiness" has a specific meaning. The clock starts only when:
- the production workflows in scope are identified;
- you've confirmed what success means for them;
- enough labeled outcomes and spend have arrived to support a comparison; and
- any sampled quality signal used has its window and coverage recorded.
It does not start at signup or at install. Getting to readiness typically takes one to three weeks after you connect, depending on your traffic. If something is missing, we tell you what, and the clock waits.
One more boundary: imported history is useful as a baseline for spend, latency, and coverage, but the brief itself comes from live runs with mapped outcomes. History is a baseline; the brief comes from live outcomes.
You get the ranked list, separate evidence for each item, an explicit "not enough evidence" section, trace links, and one next step with a verification window. You don't get a fused score, a guaranteed saving, or a cause the data can't back up.
Try it on your own agents
Morse is AI-agent observability and evals built around this brief. If you run agents in production and want a founder-reviewed brief on your own fleet, join the waitlist.