Recursive Context

How PraisonAI handles large context without context rot

Large outputs stay outside the context window — agents explore them with artifacts, delegation, and escalation instead of stuffing everything into one prompt.

from praisonaiagents import Agent, AutonomyConfig, ExecutionConfig

agent = Agent(
    instructions="You are a data analyst.",
    autonomy=AutonomyConfig(level="full_auto", max_iterations=30),
    execution=ExecutionConfig(code_execution=True),
)

agent.start("Analyse 5000 tickets and count billing issues")

The user requests analysis on huge data; the agent stores artifacts and delegates instead of overfilling context.

How It Works

Quick Start

1. Simple Usage

   Enable autonomy — artifacts, escalation, and doom-loop protection activate automatically:

   from praisonaiagents import Agent
   
   agent = Agent(
       instructions="You are a data analyst.",
       autonomy=True,
   )
   
   agent.start("Summarise the key themes in this support inbox export")
   
2. With Configuration

   Tune iteration limits, doom-loop threshold, and code execution:

   from praisonaiagents import Agent, AutonomyConfig, ExecutionConfig
   
   agent = Agent(
       instructions="You are a data analyst.",
       autonomy=AutonomyConfig(
           level="full_auto",
           max_iterations=30,
           doom_loop_threshold=3,
           auto_escalate=True,
       ),
       execution=ExecutionConfig(code_execution=True),
   )
   
   agent.start("Analyse 5000 tickets and count billing issues")
   

---

How It Works

Instead of loading entire documents into the LLM window, PraisonAI stores large outputs as artifacts and lets agents peek, grep, and chunk them on demand. When tasks grow complex, escalation increases capability stage-by-stage; sub-agent delegation handles focused subtasks with scoped budgets.

Component	Role
Artifact store	Holds large outputs; agents query with head, grep, chunk
Escalation pipeline	Progresses from direct answer → tools → plan → sub-agents
Sub-agent delegator	Spawns focused agents with token and step budgets
Doom loop tracker	Stops repeated identical actions

---

Configuration Options

AutonomyConfig

Option	Type	Default	Description
level	str	"suggest"	Autonomy level (suggest, auto_edit, full_auto)
max_iterations	int	20	Maximum turns before stopping
doom_loop_threshold	int	3	Repeated actions before doom-loop stop
auto_escalate	bool	True	Automatically increase complexity stage
mode	str	auto	caller or iterative execution mode

Artifact queueing

Large tool outputs above inline_max_bytes (default 32 KB) are stored as artifacts automatically when queueing is enabled on the agent’s tool configuration.

---

Common Patterns

Explore a large artifact

Agents with autonomy enabled can inspect stored outputs without loading the full content:

from praisonaiagents import Agent, AutonomyConfig

agent = Agent(
    instructions="Search large logs for error patterns.",
    autonomy=AutonomyConfig(level="full_auto"),
)

agent.start("Find all timeout errors in the latest deployment log")

Delegate a subtask

from praisonaiagents import Agent, AutonomyConfig

agent = Agent(
    instructions="Coordinate analysis across data sources.",
    autonomy=AutonomyConfig(level="full_auto"),
)

# Delegation runs inside the autonomy loop when escalation reaches Stage 3
agent.start("Count billing questions per region in the ticket export")

---

Best Practices

Enable autonomy for large or multi-step tasks

Set autonomy=True or AutonomyConfig(level="full_auto") when inputs exceed a few pages or require tool loops. Simple one-shot questions do not need it.

Set doom_loop_threshold for long runs

Keep doom_loop_threshold at 3–5 so agents stop when repeating the same tool call instead of burning tokens indefinitely.

Use full_auto for repository-scale work

level="full_auto" enables iterative mode, filesystem tracking, and sub-agent delegation — appropriate for code and data analysis over large corpora.

Combine with code execution for structured data

Pair ExecutionConfig(code_execution=True) with autonomy so agents can run Python over artifact contents rather than reasoning over raw text alone.

---

Related

Autonomous Loops

Run agents in iterative loops with completion signals and escalation

Context Compaction

Automatically trim context when the window fills up