# Multi-step agents and tool allowlists

Multi-step agents fail when planners pick destructive homonym tools, async approvals race writes, or memory summarization drops negation. Allowlists and human gates are coded, not hoped for. €2,400 PoC; €3,100 Eval CI; €3,900 prototype; €5,600 hardening; €8,500 MVP; audit from €8,000. Written scope arrives within 48 hours with numeric pass marks on accuracy, groundedness, p95 latency, and cost per thousand runs; deliverables include Dockerized repos, handoff notes, and Promptfoo or DeepEval fixtures your team can extend. Deliverables include Dockerized repos and Promptfoo or DeepEval fixtures wired to your CI provider.

**Sprints from €2,400 PoC · €3,100 eval CI · €3,900 prototype · €5,600 hardening · €8,500 MVP · audit from €8,000** · Topic: `ai-agents` · [Open proof](/#open-proof) · [Sample scorecard](/sample)

HTML: [/ai/agents](/ai/agents) · Markdown: [/ai/agents.md](/ai/agents.md)


_Tool allowlist and multi-step agents_

## Why teams trust this stack with Arviser

- Stack-specific gates for Multi-step agents and tool allowlists, not generic AI slideware.
- Senior engineers (10+ yrs) with daily eval tooling; founder architecture review (30+ yrs IT).
- Proof links include Open proof and sample scorecards, commercial sprints stay fixed-scope.

## Dual path

### For CTO / Staff Eng

Senior engineers (10+ yrs) build the vertical slice, wire retrieval and tools, and leave a repo your team can extend, with modular prompts, typed tool schemas, and CI eval fixtures.

### For VP Product / CFO

Fixed-hour packages (€2,400–€8,500) de-risk AI spend: kill fees on Risk-Shared PoC when narrow hypotheses fail, 100% PoC credit toward MVP on Go, and no surprise T&M overruns inside written scope.

Route: 24h PoC (€2,400) if unknown feasibility, else 40h Prototype (€3,900); add 32h Eval CI (€3,100) before scale; harden at €5,600 or ship MVP at €8,500.

## Stack-specific failure modes we gate in eval CI

- Planner selects destructive tools when descriptions share verb stems.
- Human-in-the-loop approvals time out but tool already executed async.
- Memory summarization drops negation, re-introducing banned recommendations.

### Sample eval fragment (ai-agents)

```
max_steps: 6
forbidden_tools: [deleteAccount]
```

## Which sprint fits?

| Situation | Recommendation | Sprint |
| --- | --- | --- |
| Hard unknown on this stack | 24h Technical PoC | €2,400 · 3–5 days |
| Working demo, no CI evals | 32h Eval CI harness | €3,100 · 1 week |
| Pilot traffic soon | 60h Hardening or 80h MVP | €5,600–€8,500 |

## FAQ

### Do you build Multi-step agents and tool allowlists under fixed scope?

Yes. We deliver written scope in 48h and fixed-hour packages; overruns inside agreed spec are on us.

### Can we start with eval CI only?

Yes, 32h Eval CI from €3,100 when a prototype exists but gates do not.

Reviewed by Vadim Nareyko, 30+ yrs IT · Last updated 2026-11-06

## Related

[AI hub](/ai) · [Eval CI](/services/eval-ci-harness)

