Agent anatomy: seven harness responsibilities
What eleven systems' source code reveals — and what it cannot establish
Slide contents
1. Agent anatomy: seven harness responsibilities
What eleven systems' source code reveals — and what it cannot establish
2. A leaderboard cannot explain an agent’s architecture
3. The harness turns model output into action
Agent = model + harness; evaluation wraps the agent
4. Eleven systems reveal different engineering choices
Four model-vendor systems, seven open-source, Omnigent above
5. Source code establishes structure, not effectiveness
Seven dimensions on pinned versions
6. Every system addresses all seven responsibilities
Table 1: the simplest and the fullest implementation of each subsystem
7. A small loop already does the work
Query → act → observe → repeat; the rest is operations
8. Completion can have its own check
Aider checks inside the loop, Hermes at the exit
9. Portability costs model-specific maintenance
Five strategies, from one provider to plugins
10. A tool defines an action contract
Exact replacement, patch, fuzzy cascade or shell
11. Beyond 15 tools, load descriptions on demand
From one tool to 109, and search instead of a catalog
12. Session history differs from model context
The log keeps everything; the model sees a projection
13. Working context is only one memory layer
Context compaction converged; memory writing did not
14. Permission and isolation solve different problems
May it run? · What can the process touch?
15. Delegation needs context, authority and verifiable results
Six patterns, from one agent to a cross-process protocol
16. Skills, MCP and hooks complement one another
How to work · what to connect · when to intervene
17. An absent technology is not proven useless
A corpus observation is not a causal experiment
18. A second snapshot can overturn earlier findings
Eight systems pinned in April and again in July 2026
19. The harness turns from tool into platform
SDKs, extensions, marketplaces, and a meta-harness on top
20. Claim strength depends on source and version
Source · version · measurement · interpretation
21. Start with 90 lines; grow on failure
The authors' scaffold → observed failure → new layer → measurement
22. An architecture map helps choose the test
Agent = model + a seven-subsystem harness
Most harness code serves operations, not task solving
Permission and isolation are different mechanisms
Inventories age in weeks; structure holds so far
Source observations are hypotheses for experiments
Copy the responsibility; validate the implementation on your tasks
23. The atlas suggests what experiments should test
24. Architecture needs a shared evaluation