Agenticflow LLM Models

Author: Anton Gulin · Tool: opencode-skill-creator · GitHub: @antongulin · Registry: skills.sh

AgenticFlow LLM Models

Choose the right model for your agent based on capability needs, speed requirements, and reasoning depth. Use built-in credits for all models listed here.

When NOT to use this skill

Use agenticflow-built-in-credits skill instead for pricing, credits, or billing questions. Use agenticflow-mcp skill if they need external API keys (BYOK). This skill covers model selection and capabilities, not account management or credit usage.

Orient first

af bootstrap --json

Extract models[] — this is the source of truth for available models in your workspace. Never hardcode model lists; they change between CLI releases and backend deployments. The models below are recommendations, but the live models[] array is the final authority.

Discover & health

af changelog --json           # What's new in the CLI — model additions/removals
af context --json              # AI agent orientation, env vars, invocation guidance
af bootstrap --strict --json   # Health check — exits non-zero if degraded

af bootstrap returns an invocation block and data_fresh boolean. If data_fresh: false, the backend is degraded — don't rely on stale model data from a degraded response. af bootstrap --strict exits non-zero when the backend is unhealthy, so CI/automation can abort before choosing models against a degraded workspace.

Verification rule: Before recommending any model, check models[] from af bootstrap --json. If a model is absent from that list, warn the user and fall back to a confirmed model.

Author's Top 3 Recommendations

These are the author's personal picks based on reliability, reasoning quality, and speed:

Note: deepseek-v4-flash and gemini-2.5-flash-lite may not appear in every CLI release's hardcoded KNOWN_MODELS list if they ship between releases. Always verify against af bootstrap --json > models[]. If absent, the backend may still serve them — proceed with a dry-run to confirm.

Upstream Canonical Models

The CLI's built-in validator recognizes these models as of v1.10.5. They are always safe to use:

agenticflow/deepseek-v3.2
agenticflow/gemma-4-31b-it
agenticflow/gemma-4-26b-a4b-it
agenticflow/gemini-2.0-flash
agenticflow/gpt-4o-mini
agenticflow/qwen-3.5-flash

Model characteristics

Default model change (v1.8.1+)

• Before v1.8.1: Default was agenticflow/gemini-2.0-flash
• After v1.8.1: Default is agenticflow/gpt-4o-mini
• Reason for change: Gemini 2.0 Flash refuses web_search on "latest X" prompts citing knowledge cutoff, even with explicit system prompt rules. GPT-4o-mini follows system prompts and calls tools reliably.

Model selection guide

Need a default?

# Author's primary recommendation — deepseek-v4-flash
af agent create --body '{"name":"My Agent","model":"deepseek-v4-flash","project_id":"<id>"}' --json

# Or use the upstream blueprint default — gpt-4o-mini
af agent create --body '{"name":"My Agent","model":"agenticflow/gpt-4o-mini","project_id":"<id>"}' --json

# For maximum speed with media support
af agent create --body '{"name":"My Agent","model":"gemini-2.5-flash-lite","project_id":"<id>"}' --json

Need reasoning?

• Deep verification: qwen-3.5-flash — deepest thinker
• Reliable tool use + reasoning: deepseek-v3.2 or deepseek-v4-flash

Need speed only?

• Fastest correct: gpt-4o-mini or gemma-4-31b-it

Workforce model selection

All agents in a workforce inherit the model:

# Default (v1.8.1+)
af workforce init --blueprint dev-shop --model agenticflow/gpt-4o-mini --name "My Team" --json

# Or use author's primary pick
af workforce init --blueprint dev-shop --model deepseek-v4-flash --name "My Team" --json

Reasoning configuration

Expose reasoning tokens (where supported):

af schema agent --field model_user_config --json

Check the full agent schema (for all fields):

af schema agent --json

For models with hidden reasoning (e.g. some Gemini variants), configure via thinking_config to expose reasoning tokens.

Verify model availability

# Always dry-run first
af agent create --body @agent.json --dry-run --json

The CLI validates the model string at create time. Typos fail fast with an actionable hint listing known models. If you pass a vendor/model-name-shaped string not in the known list, the CLI warns but allows it to proceed — so brand-new models work before the CLI is updated.

Avoid these models

Based on upstream changelog and known issues:

In general, if a model is absent from af bootstrap --json > models[], it may have been deprecated or renamed. Check the hint field on 400/422 errors for alternatives.

Fallback model guide

If your preferred model is unavailable:

1. Run af bootstrap --json and check models[]
2. Pick the closest match from the confirmed list above
3. Use --dry-run on create to validate before deploying
4. For reasoning-heavy tasks: fall back to qwen-3.5-flash or deepseek-v3.2
5. For speed-first tasks: fall back to gpt-4o-mini or gemma-4-31b-it

Cleanup

Test agents consume credits. Delete when done:

af agent delete --agent-id <id> --json

On errors

• 400 / Invalid model → Check models[] from bootstrap; model may have been renamed or is not yet in the CLI's hardcoded list. Try --dry-run first.
• 402 / Payment Required → Model requires credits; see agenticflow-built-in-credits skill
• 422 / Model not available → Model temporarily unavailable; the hint suggests alternatives
• finish_reason=length → Increase max_tokens in model_user_config

When hint is non-empty, follow it before retrying.