Thinking Budgets

Extended thinking token budgets for complex reasoning

Control how many tokens an agent spends on extended reasoning before producing a response.

from praisonaiagents import Agent
from praisonaiagents.thinking import ThinkingBudget

agent = Agent(
    name="DeepThinker",
    instructions="Solve problems step by step.",
)
agent.thinking_budget = ThinkingBudget.high()
agent.start("Solve this optimisation problem...")

The user asks a hard reasoning question; the agent spends extended thinking tokens up to the configured ThinkingBudget before answering.

Quick Start

1. Set a Budget Level

   Use predefined levels via the thinking_budget property after creating the agent.

   from praisonaiagents import Agent
   from praisonaiagents.thinking import ThinkingBudget
   
   agent = Agent(name="Thinker", instructions="Reason carefully.")
   agent.thinking_budget = ThinkingBudget.medium()  # 8,000 tokens
   agent.start("Explain quantum entanglement simply.")
   
2. Custom Budget

   Fine-tune token limits and adaptive scaling.

   from praisonaiagents import Agent
   from praisonaiagents.thinking import ThinkingBudget
   
   budget = ThinkingBudget(
       max_tokens=12000,
       min_tokens=1000,
       adaptive=True,
       max_time_seconds=120.0,
   )
   
   agent = Agent(name="CustomThinker", instructions="Solve complex problems.")
   agent.thinking_budget = budget
   agent.start("Design a load-balancing strategy.")
   

How It Works

Phase	What happens
1. Set	You assign a ThinkingBudget to the agent
2. Think	The model reasons up to the token limit
3. Answer	The agent returns the reasoned response

Budget Levels

Pre-configured levels for different task complexity:

Level	Tokens	Use case
minimal()	2,000	Quick answers
low()	4,000	Simple reasoning
medium()	8,000	Default balance
high()	16,000	Complex analysis
maximum()	32,000	Deep research

from praisonaiagents.thinking import ThinkingBudget

budget = ThinkingBudget.high()
tokens = budget.get_tokens_for_complexity(0.9)  # scales with complexity

CLI Usage

praisonai thinking status      # Show current budget
praisonai thinking set high    # Set budget level
praisonai thinking stats       # Show usage statistics

praisonai run --thinking

Pass --thinking <budget> on the praisonai run command to set an exact token budget for a one-off prompt:

# Use a numeric token budget
praisonai run --thinking 4096 "Explain quantum entanglement"

# Use a named level
praisonai run --thinking high "Solve this optimisation problem: ..."

# Use in --command mode
praisonai run --command my_agent.py --thinking 8000

--thinking is now correctly threaded through the direct-prompt path (praisonai run) in both --command and bare-prompt modes. Earlier releases raised a NameError on the direct-prompt path.

Value	Tokens	Notes
off	0	Disable extended thinking
minimal	2,000	Quick answers
low	4,000	Simple reasoning
medium	8,000	Balanced
high	16,000	Complex analysis
<int>	Exact	Any positive integer token budget

For persistent sessions, use praisonai thinking set <level> instead.

---

Usage Tracking

Track thinking utilisation across sessions:

from praisonaiagents.thinking import ThinkingBudget, ThinkingTracker

tracker = ThinkingTracker()
session = tracker.start_session(budget_tokens=8000)
tracker.end_session(session, tokens_used=5000, time_seconds=30.0)

summary = tracker.get_summary()
print(f"Average utilisation: {summary['average_utilization']:.1%}")

---

Best Practices

Match budget to task complexity

Use minimal() or low() for quick lookups; reserve high() or maximum() for multi-step analysis where reasoning depth matters.

Set via property, not constructor

Assign agent.thinking_budget after creating the agent — budgets are applied lazily with zero overhead when unset.

Enable adaptive scaling

Set adaptive=True so token allocation scales with task complexity via get_tokens_for_complexity().

Monitor with CLI

Use praisonai thinking stats to review utilisation before increasing budgets in production.

---

Related

Token Budgeting

Manage overall token spend across agent runs

Reflection

Self-review loops for higher-quality outputs