CrewAI boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This run examines the shipped CrewAI framework, including its memory and knowledge mechanisms in their actual callers. Its intended use is to establish how the framework produces, retains, maintains and later consumes agent context, without reducing the target to the standalone Memory API. CrewAI is an enclosing runtime: it dispatches crew tasks and owns agent model/tool execution, while also exposing standalone agents, memory and programmable Flows. The boundary is whole-system for the open-source framework at the frozen commit, not for CrewAI's separately operated hosted platform. The applicability cutoff is 2026-10-15. The overall evidence tier is code-grounded; no target execution, provider calls or tests were performed.

The source checkout's HEAD matched the acquired full commit, and its Git status was clean on inspection. All source evidence is restricted to SRC-1 at that commit. Later releases, live documentation and external repositories are not evidence for this run.

Entry points and retained-state loop

The distribution exposes the CLI through crewai_cli.cli:crewai (SRC-1, lib/crewai/pyproject.toml). The public Python exports include Agent, Crew, Task, Flow, Knowledge and a lazy Memory export (SRC-1, lib/crewai/src/crewai/__init__.py). CLI dispatch is not an excluded host: run_crew routes declarative crews, declarative Flows and classic Python projects, while the shipped crew template calls kickoff, training, replay and evaluation APIs (SRC-1, lib/cli/src/crewai_cli/run_crew.py, lib/cli/src/crewai_cli/templates/crew/main.py). These drivers belong inside the boundary.

Crew kickoff dispatches the configured process, and its task loop supplies prior outputs as context to task execution (SRC-1, lib/crewai/src/crewai/crew.py). The executor selects native tool calling or the ReAct loop and invokes the configured LLM (SRC-1, lib/crewai/src/crewai/agents/crew_agent_executor.py). These responsibilities support the runtime classification rather than a classification as a returning computation or runtime client.

if self.process == Process.sequential: result = self._run_sequential_process() elif self.process == Process.hierarchical: result = self._run_hierarchical_process() --- lib/crewai/src/crewai/crew.py

if use_native_tools: return self._invoke_loop_native_tools()

return self._invoke_loop_react() --- lib/crewai/src/crewai/agents/crew_agent_executor.py

Memory coverage includes both production and later consumption. Task execution recalls against the task description and appends formatted results to the task prompt. The executor extracts memories from the task, agent, expected result and output, then submits extracted statements for retention (SRC-1, lib/crewai/src/crewai/agent/core.py, lib/crewai/src/crewai/agents/agent_builder/base_agent_executor.py). Standalone Agent kickoff has its own recall and output-save paths in lib/crewai/src/crewai/agent/core.py; it is not treated as an external application.

query = task.description matches = unified_memory.recall(query, limit=5) --- lib/crewai/src/crewai/agent/core.py

The retention boundary includes encoding, consolidation, storage reopening, retrieval, scoped access, explicit memory tools and reset paths. Memory.remember enters the serialized save pipeline; encoding can delete, update and insert records; recall waits for pending saves and selects shallow search or RecallFlow. LanceDB storage opens an existing table when available. Agents also receive Search memory and Save to memory tools. These are shipped maintenance and consumption paths, not merely examples (SRC-1, lib/crewai/src/crewai/memory/unified_memory.py, lib/crewai/src/crewai/memory/encoding_flow.py, lib/crewai/src/crewai/memory/storage/lancedb_storage.py, lib/crewai/src/crewai/tools/memory_tools.py).

Read barrier: wait for any pending background saves to finish

so that the search sees all persisted records.

self.drain_writes() --- lib/crewai/src/crewai/memory/unified_memory.py

The boundary also includes knowledge-source ingestion and search, training artifacts consumed by later prompts, Flow state persistence and reload, and human-feedback lessons. These are distinct retained-state paths, not all interchangeable with unified memory. Knowledge assigns storage to its sources and queries that storage. Crew training saves evaluated suggestions, while agent prompt preparation applies training data. Flow runtime reloads persisted state by UUID; SQLite persistence saves and retrieves state snapshots. Human-feedback handling distills lessons into memory and recalls them to revise later outputs (SRC-1, lib/crewai/src/crewai/knowledge/knowledge.py, lib/crewai/src/crewai/crew.py, lib/crewai/src/crewai/agent/utils.py, lib/crewai/src/crewai/flow/runtime/__init__.py, lib/crewai/src/crewai/flow/persistence/sqlite.py, lib/crewai/src/crewai/flow/human_feedback.py).

Shipped prompts are included as operational evidence: the English translations contain memory extraction, query, save, consolidation and human-feedback prompts, plus the prompt slice used when recalled memories enter task context. Skill loading is also included because it admits instructions into agent context. These files are evidence of the target's contracts, not instructions to the analyst (SRC-1, lib/crewai/src/crewai/translations/en.json, lib/crewai/src/crewai/skills/loader.py).

Repository coverage

The table accounts for every top-level tracked entry. “Included” denotes the functional evidence boundary, not a claim that every file has already been read. Initial inspection traced the material entry points and state loops above; deeper implementation inspection remains available throughout SRC-1. Uninspected files are not access gaps. Tests and evaluators are included as source evidence, not as observed successful runs.

External dependencies and evidence limits

The package metadata declares model SDKs, embedding/vector-storage libraries, MCP support and optional provider integrations (SRC-1, lib/crewai/pyproject.toml). Their implementations and hosted behavior are not frozen by this repository pin. The framework's adapters, delegation boundaries and error handling remain in scope; remote model quality, service availability and provider-side storage/security guarantees do not. User-authored downstream crews, datasets, deployed services and proprietary CrewAI platform implementations are excluded. This prevents attributing host policy, application success or hosted persistence guarantees to the open-source runtime.

There is no supplied observed-run or causal-experiment evidence. Test source, evaluators and manual runners can establish intended checks and executable paths, but not passing results, reliable recall, improved task quality, latency, cost or effects of an intervention. No inspected material path was inaccessible; the principal access limit is implementation outside SRC-1. Initial sampling is not a license to describe uninspected mechanisms as established behavior.

Source register