Engineering agent
Engineering agent, planned call by call: what a coding agent asks Rhumb for this kind of app, and what the day's catalogs answer.
A local agent in Claude Code that works across GitHub, Linear, Postgres, Sentry and Slack and runs code. The developer already holds E2B and Vercel accounts, and wants a model for the agent on Anthropic or OpenAI.
1. What the agent connects to, and what it runs code on.
POST /v1/stack
{
"systems": [
"github",
"linear",
"postgres",
"sentry",
"slack"
],
"target": "local",
"harness": "claude-code",
"actions": [
"read",
"write",
"code_execution"
],
"infra": {
"vendors": [
"e2b",
"vercel"
]
}
}What must hold
- Every system named is known.
- A command-line tool is the choice wherever one is curated.
- Linear, which has no command-line tool, gets its own server.
- The sandbox is chosen among the named vendors only, with the sum behind its figure.
- No skill is offered twice.
- Every skill is a named system owner's own.
- The stack can be watched.
- With every goal stated, nothing is left to ask.
On the day's catalogs
- GitHub
- command-line tool gh
- Linear
- server app.linear/linear
- PostgreSQL
- command-line tool psql
- Sentry
- command-line tool sentry-cli
- Slack
- server https://mcp.slack.com/mcp
- Runs on
- A sandbox, because it runs code: E2B, $0.1656 for the job, the least of the vendors named, by price alone (7,200 vCPU-seconds (3,600 s × 2 vCPU) × $0.000014 = $0.1008; 14,400 GiB-seconds (3,600 s × 4 GiB) × $0.0000045 = $0.0648. Total $0.1656). Priced for a reference job at listed rates: One sandbox for an hour, with 2 vCPU and 4 GiB of memory, its processors busy the whole time.
- Warnings
- tools_not_read (linear), tools_not_read (slack)
- Skills
- sentry-create-alert (Sentry), sentry-debug-issue (Sentry), sentry-fix-stack-traces (Sentry)
2. Which model the coding agent runs on.
GET /v1/resolve?purpose=coding-agent&providers=anthropic,openai&shape=escalationWhat must hold
- A model is chosen.
- The answer is a ladder, cheapest first.
- No monthly figure is given without stated usage.
On the day's catalogs
- Step 1
- claude-haiku-5-5 on anthropic
- Step 2
- gpt-6.1-terra on openai
- Step 3
- claude-opus-5-5 on anthropic
- Reason
- a 3-step ladder for code on your providers, cheapest first: each step is the cheapest model scoring at least the next quality tier (0.16 → 0.55 → 0.95 on lmarena:webdev), on a provider no earlier step uses where possible
- Warnings
- capability_claimed_not_probed, capability_claimed_not_probed
See every use case, and Architect an agent to try a stack.