How do AI agents communicate and share state?
Inside one application, agents usually communicate through the framework: a shared state object that every step reads and writes, or messages passed when one agent hands a task to another or calls it as a tool. Across separate systems, they communicate over a protocol — A2A is the open standard for agent-to-agent communication, while MCP connects agents to tools and data. What matters most is how much context is passed: Cognition's first principle is to share full agent traces, not just individual messages.
- Published
- 22 October 2026
- Sources checked
- 22 October 2026
- From
- Kramiva
01
Within one application
LangGraph models a system as a graph whose nodes read and update shared state; a checkpointer saves that state after each step, keyed by a thread identifier, so a run can be inspected, resumed or edited by a person. The OpenAI Agents SDK passes control through hand-offs, or keeps one agent in charge and calls others as tools. Microsoft Agent Framework and Google's ADK offer comparable workflow constructs.
02
Across systems: the A2A protocol
When agents run as separate services, possibly from different vendors, they need a shared protocol. A2A — started by Google and now a Linux Foundation project — defines how agents discover each other and exchange work over JSON-RPC 2.0 on HTTP(S).
- Agent Card: a JSON description of an agent's identity, skills, endpoint and authentication
- Task: a stateful unit of work with an ID and a lifecycle
- Message and Part: a turn of communication and its content pieces
- Artifact: an output the agent produces for the task
03
What to share
Cognition's two principles are the clearest guidance here: share context, including full traces rather than just the last message; and remember that actions carry implicit decisions, so conflicting decisions produce bad results. Anthropic's research system found the same in practice — vague task descriptions from the lead agent led subagents to duplicate work or miss parts of it.
04
Trade-offs
Sharing everything keeps agents consistent but fills context windows and raises cost. Sharing summaries is cheaper but loses detail an agent may need. Cross-system protocols allow agents from different teams or vendors to work together, at the price of network calls, authentication and a larger security surface.
Where Kramiva fits
Kramiva's AI automations log every decision with what was proposed, on what basis and who approved it — the record you need when several steps or services share one task.
Sources
- LangGraph persistence — LangChain
- OpenAI Agents SDK documentation — OpenAI
- Agent2Agent (A2A) Protocol documentation — A2A Project (Linux Foundation)
- A2A key concepts — A2A Project (Linux Foundation)
- Don't Build Multi-Agents — Cognition, 12 June 2025
- How we built our multi-agent research system — Anthropic, 13 June 2025
Accessed 22 October 2026. Kramiva is not affiliated with these publishers.
Related questions
Tell us what you are building.
A short brief gets you a real reply from a founder within one business day — an honest read on fit, scope, and budget.