You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Internal Agents Map is a source-backed catalog of how organizations build and use artificial intelligence (AI) agents for internal work. It covers task agents, background agents, shared platforms, orchestration systems, and implemented supporting patterns.
The catalog collects company articles, source code, documentation, talks, social posts, news, and community commentary. A common rubric makes different approaches easier to compare. The goal is to learn from their designs, tradeoffs, results, and failures.
What the catalog includes
Organizations use the word "agent" in different ways. The catalog preserves those differences. It does not require one architecture or level of autonomy.
An approach qualifies when a named organization built or materially adapted it for its own teams, and a public source describes the implementation or its use. The approach can be an agent, workflow, platform, runtime, orchestration system, or supporting pattern. Prototypes, pilots, deployed systems, and later product releases all qualify.
Durable identity is one design choice. It is not an inclusion requirement. The comparison rubric records identity, state, invocation, autonomy, deployment stage, controls, and available evidence. Scoped operating-model assessments also show where human attention returns and derive a Shapiro level for that workflow. An unknown value means the collected sources do not document it.
Scope
The catalog includes approaches first built or materially adapted for internal work. An entry can remain internal, become open source, or become a commercial product.
Generic vendor products do not qualify without a documented internal adaptation. General opinion articles and unattributed rumors do not qualify as approaches. They can still appear as commentary when they discuss a cataloged approach.
The catalog treats company metrics as self-reported unless an independent source verifies them. It keeps source claims separate from catalog interpretation. See the data schema
for source and confidence rules. The landscape
The table is sorted by company. Each row links to the relevant section of the full catalog. The YAML records are in data/agents.
Architecture patterns compares the published technical designs. Adoption lessons collects reported operating practices. Both pages state evidence limits and distinguish company reports from catalog interpretation.
Who is this for?
Engineers who build internal agent systems
Leaders who compare implementation choices
Researchers who track reported agent designs and results
Contributors who find missing sources or conflicting evidence