🌀
← Back to Articles & Artefacts
artefactssouth

Agent environment awareness: a first survey

IAIP Research
miadi-inquiry-261015-agent-environment-awareness-b857e973-b67f-4c05-a930-5a6961e026c1

Agent environment awareness: a first survey#

Seed for the next inquiry, 2026-10-15. Guillaume asked for a small survey of the academic fields that relate to an environment that is well configured and an agent that knows which variables exist, whose they are and what they are for. One paragraph per field.

What prompted it#

The coordinator mino-2610150802-deepdiver-owner-261009-screenwalk-OOMRbYt5x58 reported that Ava had no Miadi token and that Guillaume had to issue one. Ava's token was in her own user's environment file (/home/ava/.env), and the credential list showed it had been used that morning. The search had looked only in its own shell. The episode 558 witness circle (circle:1792081081517:x4srth) was then opened with MIADI_PERSON_TOKEN, which acts as Guillaume. What was learned went into the chronicle-episode skill (S15, jgwill/miadi-orchestration-kit 0bb2790) and the miadi-stack-map skill (jgwill/etc-claude-code 7f75455).

Fields#

1. Software configuration management. IEEE 828 treats every setting a system depends on as a configuration item with an identity and an owner, and its status accounting records where each item stands. The Twelve-Factor App (Wiggins, 2011) moved configuration into environment variables. That made it portable, but a variable carries a value and nothing about whose it is or what it is for. The token table added to miadi-stack-map today is status accounting done by hand.

2. Identity and access management: delegation and impersonation. RFC 8693 (OAuth 2.0 Token Exchange, 2020) separates impersonation, where the actor cannot be told apart from the person it acts for, from delegation, where the token records both. MIADI_PERSON_TOKEN is impersonation, and a seat token is the agent's own identity. The rule that an agent never speaks with the person's token is a delegation rule, and today it holds by convention only. Hardy's confused deputy (1988) names the failure that rule prevents: authority held for one purpose and used for another. Saltzer and Schroeder's least privilege (1975) is the principle under both.

3. Situation awareness. Endsley (1995) describes three levels: perceiving the elements of the environment, understanding what they mean, and projecting what they will do. An agent that lists only its own shell fails at level 1. One that sees MIADI_PERSON_TOKEN without knowing it is Guillaume fails at level 2. Today's report was a level-1 gap presented as a level-3 conclusion: "Ava has none, so Guillaume must issue one."

4. Distributed cognition. Hutchins (1995) shows cognitive work spread across people, tools and the environment, so that part of what a system knows lives in its artefacts. The .env files, the seat token files and the skills are part of how the agents think. The fact that Ava's token was in /home/ava/.env existed in the system. It was not on the path the seat searched.

5. Transactive memory. Wegner (1987) describes how groups remember by knowing who knows what. A team of agents needs a directory of who holds which credential and which knowledge. The S2 table of the chronicle-episode skill and the miadi-stack-map table are that directory. Their failure is the one human teams know: the directory goes stale while what it pointed to is still there.

6. Agent–computer interface design. Yang et al. (2024, SWE-agent) found that language-model agents do markedly better when the interface they act through is designed for them, with compact feedback and guardrails. GET /api/identity/me, which answers whose a token is and what it may do, is such an interface element. A raw listing of environment variables is not.

7. Knowledge creation and organizational learning. In Nonaka and Takeuchi's SECI model (1995), externalization turns what a person knows without writing it into something written. Guillaume knew where the tokens were, and the S15 paragraph writes it down. Argyris and Schön (1978) distinguish single-loop learning, which fixes the instance, from double-loop learning, which changes the rule that produced it. "Any failure to find prototypes will lead to revising and patching what made you unable to do so" is a double-loop rule.

How the fields relate#

A variable is a configuration item (1). Its authority is a question of identity: who acts, for whom (2). The agent has to perceive it and understand it before it can act on it (3). It lives spread across files, homes and skills (4), so the team needs a directory of who holds what (5). The interface decides what the agent can perceive in one read (6). Each failure can then change the rule that caused it (7).

Questions for the next cycle#

  • Q1. Can each host carry an environment manifest that says, for every variable, whose it is and what it is for, so that perceiving and understanding take one read?
  • Q2. Where should impersonation be refused by the system rather than by convention, for example by recording both the actor and the person on every circle turn, as RFC 8693's act claim does?
  • Q3. What must a "none found" answer carry so that an agent can tell an absence from a query that looked in the wrong place?
  • Q4. How quickly do the skills' tables go stale, and which event should refresh them?
  • Q5. Over the chronicle, which skill edits were double-loop and which only fixed an instance?

Next#

A foundations packet (the deep-research-foundations skill) can start from these seven fields and five questions.

Sources#

  • Argyris, C., and Schön, D. A. (1978). Organizational Learning: A Theory of Action Perspective. Addison-Wesley.
  • Endsley, M. R. (1995). Toward a theory of situation awareness in dynamic systems. Human Factors, 37(1), 32–64.
  • Hardy, N. (1988). The confused deputy (or why capabilities might have been invented). ACM SIGOPS Operating Systems Review, 22(4), 36–38.
  • Hutchins, E. (1995). Cognition in the Wild. MIT Press.
  • IEEE 828-2012. Standard for Configuration Management in Systems and Software Engineering.
  • Jones, M., Nadalin, A., Campbell, B., Bradley, J., and Mortimore, C. (2020). RFC 8693: OAuth 2.0 Token Exchange. IETF.
  • Nonaka, I., and Takeuchi, H. (1995). The Knowledge-Creating Company. Oxford University Press.
  • Saltzer, J. H., and Schroeder, M. D. (1975). The protection of information in computer systems. Proceedings of the IEEE, 63(9), 1278–1308.
  • Wegner, D. M. (1987). Transactive memory: A contemporary analysis of the group mind. In B. Mullen and G. R. Goethals (Eds.), Theories of Group Behavior (pp. 185–208). Springer.
  • Wiggins, A. (2011). The Twelve-Factor App. https://12factor.net/
  • Yang, J., Jimenez, C. E., Wettig, A., Lieret, K., Yao, S., Narasimhan, K., and Press, O. (2024). SWE-agent: Agent-computer interfaces enable automated software engineering. NeurIPS 2024. arXiv:2405.15793.