No two agents doing the same job.
One agent holds each job. If it gets stuck, it hands over with notes on what it tried.
For builders
Explain it once, not to every agent.
Claude Code, Cursor, Codex and Grok read the same plan for each job (the spec). Switch agents without starting over.
What changes
You've pasted the same context into four chats this week.
Know when a job is really done.
Each plan ends with a short checklist (Done When). Every item gets checked and leaves a receipt.
Decide once. Every agent learns.
Answer an agent once and save it as a rule. Every agent reads it before asking.
Claude Code gets stuck. Cursor picks it up.
Halfway through a job, one agent stalls. It hands the job over with notes on what it tried. You open Cursor, and it reads the same spec (the plan for one job: what to build, the files to touch and when it's done) plus those notes, and carries on. Nobody re-explains anything, least of all you.
You make the call.
Annsa ranks and suggests. You pick the job, and you, your agent or Ask (Annsa's own agent) writes the spec. We build Annsa this way: 153 jobs shipped in 12 days with four agents on one list.
01
Evidence lands
Customers · roadmap · ideas · agent findings
02
Annsa ranks it
Who asked · what they said
03
You choose
The reason stays with it
04
You or your agents write the spec
Claude Code · Cursor · Codex · Grok
05
You learn what worked
Checked · decided · ruled out
The product
See the loop inside the product.
Priorities ranked in the middle, the spec they became on the right, every number tracing back to the customers who asked.
Questions people ask.
Q1
Who writes the spec?
You, your agent or Ask. Your own agent can write it while reading your repo. Ask, Annsa's own agent, can write a spec from your connected repo when you ask it to.
Q2
Can I switch agents halfway through a job?
Yes. One agent holds a job at a time. If it gets stuck, it hands over with notes on what it tried, and the next agent reads the same spec and those notes.
Q3
How does an agent get the work?
It asks Annsa for it. Each agent connects over MCP with nothing to install and signs in through your browser the first time. Start at /docs/quick-start, then /docs/connect-claude-code, /docs/connect-cursor, /docs/connect-codex and /docs/connect-grok. /docs/working-with-specs covers the rest.
Q4
Does Annsa replace my judgment?
No. Annsa ranks and suggests. You make the call, and the reason stays with the job.
Write it down once. Every agent reads the same plan.
Begin with one agent and the work that matters next.