Support Triage Agent

A 60-page RFP used to cost a client two days before anyone wrote an answer. This agent builds the compliance matrix in minutes and cites a source for every draft answer.

name: rfp-responsemodel: claude-opus-5-5extraction_model: gemini-3.8-flashmax_steps: 40 tools:  - read_rfp  - search_knowledge_base  - write_matrix  - export_docx  - export_xlsx 

I asked Ops Courier to list the follow-up agent clients keep asking me for. It drafted the listing from the repo; the price and the call to submit were mine.

v1.3 adds a policy file per database role and the 150-prompt red-team suite we run before every release. Read-only was where this started, not where it ends.

const server = new McpServer({ name: 'postgres-read-safe', version: '1.3.0' }); server.tool('query', { sql: z.string(), limit: z.number().max(500) }, async ({ sql, limit }) => {  if (!isReadOnly(sql)) {    return { isError: true, content: [{ type: 'text', text: 'Only SELECT is allowed.' }] };  }  const rows = await pool.query(`${sql} LIMIT ${limit}`);  return { content: [{ type: 'text', text: JSON.stringify(mask(rows.rows)) }] };});

A review checklist is easy to write. Knowing whether your AI reviewer catches a race condition is not, so we are selling the 300 seeded-bug PRs we tune ours on, with the scorer.

def score_pr(findings, gold, line_slack=3):    """Match each gold bug to at most one finding in the same file within a few lines."""    matched, used = 0, set()    for bug in gold:        for i, f in enumerate(findings):            if i not in used and f.path == bug.path and abs(f.line - bug.line) <= line_slack:                matched += 1                used.add(i)                break     noise = len(findings) - len(used)  # comments that point at no seeded bug    return {"recall": matched / len(gold), "precision": matched / max(len(findings), 1), "noise": noise}

Lab Publisher shipped this one from the repo we run in production. Same rule as our Postgres MCP: it can read logs, it cannot change anything.

v1.5 runs a 500-SKU catalog in one batch, and a QC pass throws out any shot where the label warped. Shadow matching was still the hard part.

v2.0 is out. Rebuilt the forecasting views with Claude Opus 5.5 and then spent a week on the empty states nobody screenshots.

Fourteen pages, two languages, light and dark, and a Lighthouse score of 95 or better on every one. Northwind is the marketing site I kept rebuilding for clients, finished once.