Every multi-agent run gets a sealed ledger: who owned what, what each agent handed back, exactly what the next agent worked from, when one stopped to ask, and whether the whole team beat doing it the simple way.
Which agent owns which task. A second agent handed the same job is flagged before it starts, or refused outright.
What sits outside its scope, sealed, and bound to a sebbi Mandate or Agent Passport when authority matters.
What it must hand back. Every output is checked against that contract and fingerprinted as it is handed back.
Each agent declares the outputs it used, by fingerprint. A verified handoff proves it worked from exactly what it was given.
When it must stop or ask a human. Every escalation is sealed with its reason.
The run report sets the team against a simpler approach you name: wall time, agent time, how much ran in parallel, how long agents sat waiting for each other, tasks that duplicated each other, tokens and cost. The verdict is sealed with the report.
# 1. Start a run, with the simpler approach to beat
curl -X POST https://sebbi.pro/x/agentrun/start -H "Authorization: Bearer $SEBBI_KEY" \
-d '{"name":"Competitor research","baseline":{"minutes":40,"cost_gbp":1.20,"note":"one agent, sequential"}}'
# 2. Give each agent one job, its limits, what it must return and when to stop
curl -X POST https://sebbi.pro/x/agentrun/task -H "Authorization: Bearer $SEBBI_KEY" -d '{
"run":"RN-...","task_id":"pricing","agent":"researcher-1",
"objective":"Collect published UK pricing for the five named competitors",
"boundaries":["public web only","no logins","no emails"],
"returns":{"fields":["competitors","sources"],"max_chars":20000},
"escalate_if":["a source is behind a paywall","fewer than 3 competitors found"]}'
# 3. Seal what it hands back
curl -X POST https://sebbi.pro/x/agentrun/output -H "Authorization: Bearer $SEBBI_KEY" \
-d '{"run":"RN-...","task":"pricing","output":{"competitors":[...],"sources":[...]},"tokens":18000,"cost_gbp":0.22}'
# 4. The next agent proves exactly what it worked from
curl -X POST https://sebbi.pro/x/agentrun/output -H "Authorization: Bearer $SEBBI_KEY" \
-d '{"run":"RN-...","task":"summary","output":{...},"used_inputs":[{"task":"pricing","sha256":"<hash from step 3>"}]}'
# 5. Stop or ask a human, sealed. Then finish: the report is sealed too.
curl -X POST https://sebbi.pro/x/agentrun/escalate -H "Authorization: Bearer $SEBBI_KEY" -d '{"run":"RN-...","task":"pricing","reason":"paywall"}'
curl -X POST https://sebbi.pro/x/agentrun/finish -H "Authorization: Bearer $SEBBI_KEY" -d '{"run":"RN-..."}'
Outputs never leave your systems: sebbi.pro keeps fingerprints, field names and figures. Each public report shows hashes and numbers only, at https://sebbi.pro/run/<run id>, and anyone can check a handoff against it. 50p per agent per month after 7 days free. Works with any framework and any model. Pairs with the Mandate Chain for provable authority between agents.