Fail-Loud Defaults

PraisonAI raises clear errors instead of silently picking a default when configuration is ambiguous.

from praisonaiagents import Agent

# Explicit provider/model — no silent OpenAI fallback
agent = Agent(name="assistant", llm="ollama/llama3")
agent.start("Hello!")

The user passes ambiguous configuration; PraisonAI raises an explicit error with a fix hint instead of silently falling back.

How It Works

Behaviour change (PR #2122): Several subsystems that previously fell back silently now raise explicit errors. Review the table below when upgrading.

Also fixed in PR #4128: Seven Agent(...) preset params — output, execution, context, autonomy, approval, learn, self_improve — now raise ValueError on a typo instead of silently using defaults. The worst case was security-relevant: approval="read_onl" produced an empty deny set, leaving all 17 dangerous tools callable. If you rely on the previous silent-default behaviour, pass the default explicitly: approval=False, execution="balanced", etc.Completed by PraisonAI PR #4186 (issue #4182): the same four params also silently disabled themselves when the validator accepted a spelling the resolver did not (autonomy="full-auto", context="sliding-window", "FULL_AUTO", " thorough ", …). Validation and resolution now share one canonical spelling rule (canonical_preset_key() in praisonaiagents/config/parse_utils.py), so any spelling the validator accepts also resolves to the intended preset. approval, learn, and self_improve were already correct; this closes the gap for the remaining four (output, execution, context, autonomy).

A broken optional dependency that failed with anything other than ImportError used to re-probe on every hot-path access; now the first failure is cached and re-raised fast (PR #4205).

What Changed

Area	Old behaviour	New behaviour	Deep dive
Database URL	Unknown scheme → SQLite fallback	`ValueError`	Cloud Databases
Model string	Unrecognised name → OpenAI	`ValueError`	Multi-Provider Advanced
Daytona sandbox	Appeared available with client	`NotImplementedError`	Sandbox
LazyCache	Only `ImportError` cached; other loader failures re-ran on every call	Caches and re-raises every terminal loader failure (`ImportError`, `RuntimeError`, `OSError`, …); `KeyboardInterrupt` / `SystemExit` still propagate uncached — PraisonAI PR #4205	Optional deps docs
Approval	`enabled: false` by default	`enabled: true` by default	Approval
Claude CLI backend	`bypassPermissions` default	`default` mode; bypass requires opt-in	CLI Backend Protocol
SandlockSandbox (PR #1367)	Landlock ABI too low → silent `SubprocessSandbox` fallback	`RuntimeError` at init time	Sandbox
`Agent()` preset params (PR #4128) — `output`, `execution`, `context`, `autonomy`, `approval`, `learn`, `self_improve`	Typo → silent fallback to defaults (worst case: `approval="read_onl"` → empty deny set, all tools callable)	`ValueError` with “Did you mean …?” suggestion	Agent Presets & Modes, Approval

Quick Start

Grep your logs for new errors

grep -r "Unable to infer DB backend\|Cannot infer provider\|Daytona backend not yet implemented" logs/

Fix database URLs explicitly

# Was silently SQLite — now raises ValueError
db_url = "sqlite:///mydata.db"  # explicit scheme required

Use provider/model form

from praisonaiagents import Agent

# Was OpenAI fallback — now raises ValueError
agent = Agent(name="assistant", llm="ollama/llama3")

Migrating

Exception text	Fix
`Unable to infer DB backend from URL '...'; supported schemes: postgres://, mysql://, sqlite://, redis://, libsql://, http(s)://`	Prefix URL with a supported scheme, e.g. `sqlite:///path.db`
`Cannot infer provider from model '...'. Use the 'provider/model' form, e.g. 'ollama/llama3', 'bedrock/anthropic.claude-3-sonnet'.`	Use `provider/model` or a recognised prefix (`gpt-`, `claude-`, `gemini-`)
`Daytona backend not yet implemented. Use 'subprocess', 'docker', or 'e2b' sandbox instead.`	Switch `sandbox_type` to a supported backend
Approval prompts where none expected	Set `approval=False` or `approval: {enabled: false}` in YAML
`Unknown preset '<typo>' for parameter '<param>'. Did you mean '<suggestion>'?`	Fix the spelling, or pass the default explicitly (`approval=False`, `output="silent"`, `execution="balanced"`, …). Free-form params (`knowledge=`, `guardrails=`, `skills=`, `runtime=`, `auth=`) are exempt from this check.
Silent fallback to a disabled feature after passing a hyphen / padded / uppercase preset (e.g. `autonomy="full-auto"` behaves as `autonomy=False`)	Upgrade to a build including PraisonAI PR #4186. From that build on, `full-auto`, `FULL_AUTO`, and `full_auto` all resolve to the same preset as `full_auto`.
A cached non-`ImportError` (e.g. `RuntimeError` from a broken native library) now re-raises on every access to an optional feature	Install the missing system dep or set the feature off explicitly — the cache no longer masks the failure by re-running the loader each call

Best Practices

Prefer explicit configuration

When in doubt, spell out schemes, provider prefixes, and backend names — silent fallbacks are gone.

Disable approval only when intended

Approval is on by default (PR #2122). Use approval=False for fully autonomous runs.

Never use BYPASS without env gate

Claude Code bypass requires unsafe=True and PRAISONAI_CLAUDE_BYPASS_PERMISSIONS=1.

Pin sandbox backends explicitly

Do not rely on implicit fallbacks — set sandbox_type to a supported backend (subprocess, docker, or e2b).

Approval

Safe-by-default approval gates

Cloud Databases

Supported DB URL schemes

Getting Started

Install

Core Concepts

Guides

Features

Models

Databases

Observability

Memory

Telemetry & Analytics

Knowledge

RAG

Persistence

Tools

Other Features

Developers

Configuration

Best Practices

Getting Started (No Code)

Fail-Loud Defaults

How It Works

What Changed

Quick Start

Migrating

Best Practices

Approval

Cloud Databases

​How It Works

​What Changed

​Quick Start

​Migrating

​Best Practices

​Related

Approval

Cloud Databases

How It Works

What Changed

Quick Start

Migrating

Best Practices

Related