Can loose office habits and shared templates let AI agents coordinate as well as a strict formal communication protocol?
Can informal codified behavior compete effectively with formally standardized agent communication?
This explores whether agents coordinating through codified conventions, such as standard operating procedures, shared document formats and learned social norms, can hold their own against formal, standardized communication protocols, or whether the two do different jobs.
This explores whether agents that coordinate through codified habits (workplace procedures, shared templates, social norms) can match agents that talk through formal, standardized protocols. The corpus has no head-to-head comparison. What it does suggest is less obvious: the line between 'informal' and 'formal' is blurrier than the question implies, and the two seem to win in different places.
The strongest evidence that informal practice can compete comes from an unexpected source: ordinary office habits. MetaGPT took human standard operating procedures, the unwritten routines software teams follow, and turned them into structured documents that agents produce and pull from a shared workspace. That beat agents simply chatting with each other Does structured artifact sharing outperform conversational coordination?. The lesson isn't that formal protocols beat informal behavior. It's that informal behavior becomes powerful once you write it down in a fixed shape. A related idea is that code itself can be the shared medium: it can run, be inspected, and hold state, so it serves as both the 'convention' and the 'contract' at once Can code serve as the operational substrate for agent reasoning?.
At the far informal end, there's a proposal for agents that discover and follow social norms on the fly, using world knowledge and theory-of-mind reasoning instead of hard-coded rules Can LLMs discover and apply social norms dynamically in open systems?. It's attractive for open settings where nobody wrote the rulebook in advance, but the authors admit no implementation has been validated. So the most flexible version of informal coordination is also the least proven.
Formal standards earn their place at the edges between systems, especially around accountability. Chan et al. argue that training good behavior isn't enough: you also need outside infrastructure that records which agent did what, shapes how agents interact, and catches and repairs harm What infrastructure must exist beyond training agents?. Norms that live inside an agent's head can't provide an audit trail. OpenAI's 'scope of autonomy' records point the same way: they write down allowed tools, spending limits and forbidden actions, rather than trusting the agent's judgment How can autonomous agents act safely within defined boundaries?. Even inside a single agent, reliability comes from moving protocols out of the model and into a surrounding harness Where does agent reliability actually come from?.
The takeaway you might not have expected: winning formal standards don't stamp out existing conventions. They wrap them. Coordination layers get adopted by bridging protocols that already exist, like MCP, not by replacing them Should coordination protocols wrap existing systems or replace them?. That points to a layered picture rather than a contest. Codified informal practice handles coordination inside a team. Formal standards handle trust, identity and accountability across organizations. And one research thread sidesteps the debate entirely: agents sharing internal 'thoughts' directly, without language, which can reveal disagreements before they surface in words Can agents share thoughts directly without using language?.
Sources 8 notes
MetaGPT demonstrates that agents producing standardized engineering documents achieve superior coordination compared to conversational exchange. Active information pulling from shared environments eliminates noise and mirrors efficient human workplace infrastructure.
Research shows code uniquely enables agent reasoning, action, and verification by being simultaneously executable, inspectable, and stateful. This unified code-centered loop improves reasoning and verification together compared to natural-language or prose-based approaches.
A plausible design exists to extend LLM-agent modules with norm discovery, reasoning, and conformance by leveraging existing capabilities like world knowledge and Theory-of-Mind reasoning, though no implementation has been validated.
Chan et al. argue that external technical systems modeled on internet protocols are essential beyond behavior training. These systems must handle three core functions: attributing actions to agents, shaping their interactions, and detecting and remedying harm.
OpenAI's Model Spec formalizes agent autonomy through structured scope records listing allowed tools, cost limits, and prohibited high-risk activities, mirroring software security's principle of least privilege while avoiding confirmation overload.
Show all 8 sources
Research shows reliable LLM agents externalize three cognitive burdens—memory (state persistence), skills (procedural components), and protocols (structured interaction)—into a harness layer rather than relying on model scale alone. The harness unifies these externalities and eliminates the need for the model to solve the same problems repeatedly.
Research shows that agent coordination standards achieve adoption by composing existing protocols like MCP and DIDComm under a shared substrate, rather than competing to replace them. Bridging lets value accrue incrementally without forcing ecosystem-wide rewrites.
Research formalizes inter-agent thought sharing via sparse autoencoders that recover individual, shared, and private latent thoughts from hidden states. This approach detects alignment conflicts at the representational level before they manifest in language.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Toward an Organizational Science of Multi-Agent LLM Systems: Decoupling Who, How, and Which Algorithm
- Code as Agent Harness
- Harness Updating Is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents
- MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation
- Infrastructure for AI Agents
- Scaling Behavior of Single LLM-Driven Multi-Agent Systems
- Embodied LLM Agents Learn to Cooperate in Organized Teams
- Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures