Mistral released Vibe 2.0 on January 27, 2026, highlighting custom subagents, multiple-choice clarification, slash-command skills, and unified agent modes. Vendor demonstrations and a few trials are not universal production evidence. My view is that subagents do not automatically improve reliability. They help when task contracts, permissions, and verification become clearer.

A subagent is an interface first

A “code review agent” should not be merely another long prompt. It needs explicit input: change range, repository standards, base revision, and readable files. It needs explicit output: findings ordered by severity, evidence locations, and uncertainties. It needs an authority boundary: read-only, with no file editing or deployment. If its assignment is “see whether anything is wrong,” it inherits all ambiguity from the coordinator.

A small workflow might define three roles. An explorer gathers relevant files and constraints. An implementer edits only the approved scope. A verifier runs named checks and explains failures. They exchange compact, structured handoffs rather than unlimited chat histories. Each completion reports inspected material, conclusions, unresolved risks, and a proposed next action; the coordinator decides whether to continue.

Keep one writer for overlapping files. Parallel exploration is usually safe, but several agents editing the same module create conflicts and duplicate work. More concurrency is not progress if the user must reconcile it manually.

Clarification can be a reliability mechanism

Vibe 2.0’s multiple-choice clarification matters because many dangerous terminal-agent failures begin with a reasonable but wrong interpretation. Ask where an answer changes implementation or side effects: replace or merge configuration, patch the bug or refactor the module, install a dependency or remain within the lock file. Repository-discoverable naming and formatting should not interrupt the user repeatedly.

Options should be mutually exclusive and state their impact. Record the decision in task state so each subagent uses the same assumption. If a read-only inspection can establish the answer, inspect first. Clarification must not become a way to avoid basic discovery.

Measure clarification quality as well as frequency. A workflow that asks ten low-value questions may avoid mistakes by transferring all analysis to the user. The useful question is the smallest one that resolves a material branch.

Skills should make workflows reproducible

Slash-command skills fit repeated, reviewable procedures such as a release check, schema-migration audit, or dependency update. A skill should specify inputs, command sequence, stop conditions, and evidence. It should not hide an unlimited shell template. Version skills with the repository and review changes like code.

Never embed credentials, personal paths, or production hosts. Validate explicit parameters. Default to read-only. Ask for human approval immediately before deployment, deletion, or permission change. Tool output can include malicious instructions from a website or dependency; treat it as data, not as authority to expand permissions.

Skills also need failure contracts. A failed formatter may be repairable; a failed migration checksum should stop the workflow. Continuing after every nonzero exit merely produces a polished but untrustworthy final report.

Modes should describe authority, not personality

“Fast” and “careful” are difficult to verify. “Read-only diagnosis,” “editing without installation,” and “tests allowed” are concrete. A mode can enumerate tools, writable directories, network access, and actions that require approval. Log changes and keep the current authority visible.

A subagent should inherit or narrow its parent’s permissions, never gain authority through delegation. A deployment specialist can prepare a plan without receiving a production credential until the user approves the exact deployment. Isolation is more valuable than a role name.

Compare workflows on the same tasks

Create recoverable repository tasks: a one-file defect, a multi-module change, an ambiguous request, a failing test, and a README containing an adversarial instruction. Compare a single agent with the subagent workflow on task success, unrelated edits, verification evidence, clarification count, tokens, elapsed time, and human corrections. Repeat configurations; one successful demonstration proves little.

Classify failures: missed standard, wrong edit, missing validation, permission overreach, delegation loop, or lost context. A workflow may improve completion but cost too much latency for routine work. A read-only reviewer may be useful even if a multi-agent implementation is not.

Mistral’s announcement establishes what Vibe 2.0 offers; claims about reliability remain application-specific inference until measured. That separation matters when evaluating any coding-agent product.

Custom subagents, clarification, skills, and modes provide better orchestration primitives. Reliability comes from narrow contracts, least authority, one writer, and repeatable evaluation—not a larger cast of agents. Before adopting a role, require it to state input, output, permissions, and completion evidence. If those boundaries cannot be written, delegation is likely to add uncertainty rather than remove it.