Let Jev run the handoffs
An agent team makes the same small calls all day: is this worth reading, which agent takes it, does the claim hold, have we seen it before. A flow answers each one with a Jev judgment in under half a second, so expensive models only see the work that needs them.
Watch every item find its lane
Gates, desks and lanes on one live graph, with the run log, pass rates, the ledger and what Jev cost. This room replays invented items; the tester at the bottom runs real ones on the live flow.
Run log
click a row to light up its routeLedger · 12h
- Screened
- 0
- Sent to a desk
- 0
- Notes checked
- 0
- Duplicates
- 0
- To the frontier LLM
- 0
- In your brief
- 0
Items per hour
Run a real item through the Research desk
The route this item takes shows up here, and lights up on the graph.
Four parts, no glue code
Evally decides where each item goes. Your agents, queues and scripts do the work the lane names.
Gates
Each stage is a judge: a yes/no, a choice between desks, a score. Every verdict has a route, and anything unexpected takes the fallback.
Lanes
Where an item ends up: an agent's queue, a delivery like the morning brief, storage, an escalation to a person or a frontier model, or the bin.
Entries
Work comes back in. When a desk agent writes its note, the note re-enters at the fact-check, then the dedupe, then the brief.
Control room
A live graph of where items go, the run log, each gate's pass rate, items per hour and what Jev cost, at $0.042 per million input tokens.
Start from a team that already works
Every stage is one of the ready-made judges, so each flow runs as-is. Fork one, then point any stage at your own judge.
Research desk
Items from arXiv, GitHub, RSS and Hacker News come in. Jev skips what isn't worth reading and sends the rest to the right desk agent. When a desk hands back its note, Jev checks the claim against the source, drops duplicates, and decides what makes the morning brief. Only the unsure go to a frontier model.
Support inbox
Jev filters spam, then routes each real ticket to the agent that should handle it: on-call, billing, bugs, the help-center bot or sales. When an agent drafts a reply, it comes back in for review: good replies go out, the rest go to a person.
Agent tool-call gate
Every request is checked for injected instructions first. Clean requests go on to the tool-call firewall, which allows safe calls, asks the user before risky ones, and denies anything the task didn't ask for.
Content pipeline
Drafts from your writing agent are checked for AI slop and brand voice before anyone reads them. Clean drafts get a virality score: the strong ones publish, the rest go back to the rewrite agent with the reason.
One call per item. The lane comes back.
Send an item in; get the lane to hand it to and every gate it passed. Each gate is one judgment from your plan, 500 free a month.
curl -X POST https://tryevally.com/api/v1/flows/research-desk/run \ -H "Authorization: Bearer $EVALLY_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "entry": "intake", "input": { "title": "A benchmark for tool-use reliability in LLM agents", "text": "We measure how often agents call tools with malformed arguments.", "source": "arxiv" } }'{ "entry": "intake", "lane": { "key": "papers", "label": "Papers desk", "kind": "agent" }, "path": [ { "stage": "screen", "verdict": "read", "next": "route", "ms": 507, "signal": { "q": "relevance", "type": "score", "value": 0.91 } }, { "stage": "route", "verdict": "papers", "next": "papers", "ms": 345, "signal": { "q": "desk", "type": "choice", "choice": "papers", "value": 1 } } ], "ms": 852, "usage": { "input_tokens": 877, "output_tokens": 113 }}Fork the Research desk and point it at your feeds
Free for 500 judgments a month. No credit card.