Agent teardown: infer a counterpart agent's configuration from its messages #123

New issue

Closed

opened 2026-08-11 23:40:56 +00:00 by coilyco-ops · 0 comments

coilyco-ops commented

2026-08-11 23:40:56 +00:00

Member

Outcome

Sirens Deep can read another agent's messages in a shared channel and say what they reveal about it: memory model, context bounds, tool surface, and where its boundaries sit.

Why this target and not a storefront

The original sketch pointed Deep's read-only Steam MCP at store pages, looking for pricing games and dark patterns. The lens is right and the target is dull.

The lens comes from Abhay Kashyap's teardown, which audits a website end to end for architecture, API surface, trackers, and dark patterns through a three-agent pipeline of investigator, analyst, and editor, strictly read-only. Abhay is on the August 19 panel, so the reference is legible to him without explanation.

Aim that same lens at the subject the show is actually about. Agents reading agents, in a room full of them, on a stream premised on agents meeting for the first time.

It also needs no new tool. Deep reads messages in a channel it is already in.

This is also the eval

Agent-to-agent recognition is the hardest axis in #81, the one #81 says makes the eval interesting, and the one #76 records as unbuilt. Recognizing a counterpart as an agent rather than a person is the first step of describing what kind of agent it is. This feature and that axis are one build, not two.

Shape

Bounded to what is observable in the channel. No probing, no eliciting configuration through manipulation, no attempts to get another agent to exceed its own boundaries.
Observation and inference stay distinguishable in the output. "It answered without being mentioned" is observation. "It probably has no mention gate" is inference. Collapsing the two is the failure mode.
teardown's hard rules are the right model for tone as well as authority: read-only, no modification, no targeting anyone beyond the participants who opted into the demo.

Boundaries

Every counterpart in scope belongs to a panelist who agreed to the shared-Discord experiment. No bystanders.
Public-safe on a permanent recording, which means no speculation about a person from their agent's behavior.

Depends on

coilyco-bridge/deploy#365 for guild access. Deep currently refuses every guild summon by design.

## Outcome Sirens Deep can read another agent's messages in a shared channel and say what they reveal about it: memory model, context bounds, tool surface, and where its boundaries sit. ## Why this target and not a storefront The original sketch pointed Deep's read-only Steam MCP at store pages, looking for pricing games and dark patterns. The lens is right and the target is dull. The lens comes from Abhay Kashyap's [teardown](https://github.com/hayabhay/teardown), which audits a website end to end for architecture, API surface, trackers, and dark patterns through a three-agent pipeline of investigator, analyst, and editor, strictly read-only. Abhay is on the August 19 panel, so the reference is legible to him without explanation. Aim that same lens at the subject the show is actually about. Agents reading agents, in a room full of them, on a stream premised on agents meeting for the first time. It also needs no new tool. Deep reads messages in a channel it is already in. ## This is also the eval Agent-to-agent recognition is the hardest axis in #81, the one #81 says makes the eval interesting, and the one #76 records as unbuilt. Recognizing a counterpart as an agent rather than a person is the first step of describing what kind of agent it is. This feature and that axis are one build, not two. ## Shape - Bounded to what is observable in the channel. No probing, no eliciting configuration through manipulation, no attempts to get another agent to exceed its own boundaries. - Observation and inference stay distinguishable in the output. "It answered without being mentioned" is observation. "It probably has no mention gate" is inference. Collapsing the two is the failure mode. - teardown's hard rules are the right model for tone as well as authority: read-only, no modification, no targeting anyone beyond the participants who opted into the demo. ## Boundaries - Every counterpart in scope belongs to a panelist who agreed to the shared-Discord experiment. No bystanders. - Public-safe on a permanent recording, which means no speculation about a person from their agent's behavior. ## Depends on `coilyco-bridge/deploy#365` for guild access. Deep currently refuses every guild summon by design.

coilysiren closed this issue

2026-08-12 02:47:39 +00:00

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

coilyco-gaming/sirens-echo#123

No description provided.

Rows
Columns