Context Windows have grown large. Agents have started running beyond context window limitations.
This has created a new set of problems: how to prevent your agent from losing the plot. I call this problem Agent Drift.
More on it here
https://t.co/S23JWYMHtN
wanted my claude agents to speak a2a. didn't want to write the protocol plumbing.
built fastharness:
@harness.agentloop(
name="researcher",
description="Deep research agent",
)
async def researcher(prompt, ctx, client):
result = await https://t.co/Uh85qCzOL8(prompt)
while needs_more(result):
result = await https://t.co/Uh85qCzOL8("dig deeper")
return result
custom control flow, multi-turn reasoning and it's a2a compliant out of the box. FastAPI under and over the hood, works well with your existing middleware
https://t.co/wyDpCsvE5q
wrote a builder's guide to a2a here: https://t.co/FYayBb3ZHZ
covers agent cards, discovery, auth, and the four pillars of the agent economy. plus working python examples.
mcp is how an agent uses a tool.
a2a is how agents talk to each other.
we keep conflating them. they're not competitors. mcp is hands, a2a is language.
one without the other and you've got agents that can act but can't coordinate. or agents that can coordinate but can't do anything.
built a mesh network for ai coding agents.
two claude code sessions in different tmux windows can now talk to each other. frontend asks backend "what endpoints do you have?" and gets a real answer.
agents coordinating without a human in the loop. feels like the future