A thousand sessions.
One person’s attention.
You never watch a thousand sessions. You write one prompt with variables, prove it on three items, let Frazer run the rest under hard budgets and lane caps, and review clusters of similar results. Your attention goes to the dry run, to questions asked once for many sessions, and to the results that differ.
- A
- Seven buckets, not 1,000 rowsQueued, running, waiting on you, needs review, done, failed and skipped, with cost against the cap and an ETA.
- B
- Clusters by the shape of the changeResults grouped by diff, tests and the verifier’s verdict; outliers stand apart.
- C
- Asked once, answered onceQuestions with the same meaning merge into one; your answer goes to all of them and becomes a standing answer.
- D
- Limits you can seeConcurrency, the run’s model key and the Agent PR lane, with the kill switch armed.
One prompt, in waves.
Each wave earns the next.
A fan-out is a personal campaign: your prompt, your items, your budget. Being wrong costs three items, not a thousand, because nothing else starts until you accept the dry run.
The defaults
- Items
- Up to 5,000 per run, from a catalog query, repositories, tickets, a dataset or a list
- Concurrency
- 100 sessions at a time per person by default; an org admin may raise it to 1,000. Project quota still applies
- Waves
- A dry run on 3, then 10 %, then the rest. The next wave starts when you accept the dry run, then while guardrails hold
- Budget
- Per item and per run, reserved before each model call; an alert at 80 %, a hard stop at 100 %
- Pull requests
- Drafts only, one per repository, 40 an hour by default, on the Agent PR lane
- Kill switch
- For the run, a wave or an item. Trips above 25 % failed after 20 items, on a guardrail breach, at the budget, or after 30 minutes without progress
- Who works
- The Frazer agent, one headless session per item; or a squad, whose manager hands items to up to 12 workers
apiVersion: frazer.ai/v1
kind: Campaign
metadata: {name: pgx-v57, project: acme-staging}
spec:
owner: user:lena # a person, or a squad
items:
from: catalog
query: kind=module AND language=go AND deps.pgx<5.7
prompt:
template: |
Upgrade {{module}} to pgx v5.7 and keep its public API.
Run {{test_cmd}} in {{path}}; commit on {{branch}}.
run:
harness: frazer-agent
image: {tier: standard, class: std-2}
route: balanced # priority class: batch
concurrency: 100
perItem: {budgetUsd: 1.50, timeout: 45m}
budgetUsd: 400 # alert 80 %, stop 100 %
lane: agent-pr
waves:
- {name: dry-run, items: 3, gate: person}
- {name: wave-1, percent: 10, gate: auto}
- {name: rest, percent: 100, gate: auto, prRate: 40/h}
outcomes: {cluster: true, verifier: frazer-reviewer}
notify: digestReal machines for every session.
None of them your laptop.
Your own sessions run on your devbox. A fan-out’s sessions run headless, each in its own microVM started from a warm pool, so a thousand of them never touch your box, and never slow the session in front of you.
Agent images
Fan-out items run on Lite or Standard; Full and GPU need a plan. Skills, tools, Knowledge and models are an overlay applied when a session starts, so changing them needs no rebuild.
Results come back as decisions.
Not as a thousand pings.
Before anything reaches you, the Frazer Reviewer checks each result at its exact commit and a verifier from another model family passes or rejects it. What is left is grouped, so the effort follows the number of different changes, not the number of items.
One row per run of work
Agent activity collapses into a single row, “Worked 9 min · 4 commands, 6 edits”, that opens to one-line steps. Never paragraphs of agent chatter.
A few interrupts, then a digest
Item events never notify. You hear about the dry run, a batched question, a tripped kill switch or 80 % of budget, and the end of the run; the rest waits for the digest.
Review by cluster
A representative, the median diff, and three random spot checks, then accept or reject the whole cluster. A member that differs leaves as an outlier, reviewed one by one.
Questions asked once
When 174 sessions ask the same thing, it is one card. Your answer goes to all of them, and “make it a standing answer” covers the sessions that have not asked yet.
Designed for 1,000.
Here is what stands behind it.
Frazer is designed for 1,000+ concurrent sessions per organization. That is design capacity, not a measurement. What stands behind it is below: the default quotas that keep a thousand sessions manageable, and what every run reports.
- Concurrent fan-out sessions
- 100 per person by default · an org admin may raise it to 1,000
- Items in one run
- up to 5,000
- Agents per person
- 25 by default · the organization may raise it to 1,000
- Workers per squad manager
- up to 12
- Sub-agents
- two levels deep, four children per run
- Pull requests from a run
- drafts, one per repository, 40 an hour by default
- Agent PR lane
- below people’s pull requests · 4-runner floor · 40 % cap · a daily budget that holds jobs
- Kill switch
- above 25 % failed after 20 items · a guardrail breach · the budget · 30 minutes without progress
How a run is measured
In every run’s report, from the run’s own records.
- Person-minutes per 100 accepted results, the measure that matters
- Accept rate by cluster, and cost per accepted result
- Item time p50 and p90, queue wait, items per hour
- Rate-limit backoff and Agent PR lane hold time
- Reviewer findings per accepted PR, reverts within 7 days
How your team measures the impact
On your own records: your repositories, CI, deploys and incidents.
- Cycle time and pull requests merged, split by agent and human
- Time to merge, from the pull request to main
- Deployment frequency and change lead time, from Frazer deploys
- Change fail rate and recovery time, from incidents and rollbacks
We publish no percentages for these until we have measured them, and then with the method.
Keep going.
How it works
One piece of work, from a prompt to production, in eight steps on the real screens.
Stay in charge
Approvals, budgets, guardrails, grants and Secrets, audit and the kill switch.
Tour
Twelve screens of the Frazer app, and every way into the ADE.