THE AGENT HARNESS · SINCE 1 JULY 2026

131 DOORSACROSS 17 PLANES

An aircraft maker knows every part of the plane. We know every moving part of an AI agent: each place it can be turned against the people it works for is named, numbered, and checked in the path of the action.

131
doors
17
planes
42
doors on 1 July 2026

The registry, growing · since 1 July 2026

01 JUL 2026
42
doors · 8 planes
08 JUL 2026
48
doors · 9 planes
18 AUG 2026
68
doors · 11 planes
15 SEP 2026
112
doors · 16 planes
16 SEP 2026
114
doors · 16 planes
27 SEP 2026
130
doors · 17 planes
22 OCT 2026
131
doors · 17 planes

The 17 planes · click one

1LAUNCH & CONFIG92TOOLS & CAPABILITIES63CONTEXT & DATA94MODEL & REASONING65OUTPUT & ACTION96MULTI-AGENT97IDENTITY & AUTHORITY98RUNTIME · INFRA · SUPPLY · TIME149TRANSPORT710BREAKOUT411CHAINGRAPH1712MACHINERY1013SILICON114HUMAN OVERSIGHT415ECONOMY816LIFECYCLE & LAW417THE UPDATE CHANNEL5

Why doors, not a filter

Everything that moves is a way in

A package that updates, a token that refreshes, a setting a repository hands the runner, a second agent in the chat. If it changes and it touches the agent, it will be used.

Checked in the path

A door stands where the action happens. The action does not run until the door holds. Nothing is inferred from what the model says it will do.

Fails closed

A door with no evidence does not open. Evidence the agent reports about itself does not count. Silence is never a pass.

PLANE 1

LAUNCH & CONFIG

9 doors · 1–7, 115, 121

Everything read before the first prompt: config, hooks, the runner, the settings a repository hands it.

1Tool / MCP manifest (mcp.json)
2Startup / lifecycle hooks
3System prompt / instruction file
4Agent config (model/allowlist/perms)
5Environment vars / secret injection
6Sampling params (temp/top-p/logit-bias)
7Feature / runtime flags
115Harness plumbing
121Launch surface beyond git
PLANE 2

TOOLS & CAPABILITIES

6 doors · 8–11, 83, 130

The tools and the browser the agent acts through.

8Tool definitions & schemas
9Tool registry / dynamic discovery
10Tool responses (poisoned returns)
11Structured-output / function-call schema
83Rendered-surface provenance
130Browser hygiene
PLANE 3

CONTEXT & DATA

9 doors · 12–17, 84, 97, 129

What the agent reads: memory, retrieval, caches, hidden text.

12Prompt inputs (user / other agent)
13RAG / retrieved documents
14Memory (short & long term)
15Files the agent reads
16Embeddings / vector store
17Conversation history / context window
84Hidden-character injection
97Tenancy isolation
129Cache provenance
PLANE 4

MODEL & REASONING

6 doors · 18–21, 98, 105

The model itself: the weights, the evaluation, the grader.

18Model weights / registry
19Fine-tune / adapter / LoRA
20Chain-of-thought / reasoning trace
21Eval / probe behaviour (sandbagging)
98Eval-to-prod binding
105Evaluator independence
PLANE 5

OUTPUT & ACTION

9 doors · 22–26, 99, 101, 110, 122

What the agent does: writes, sends, deletes, pays.

22Action path / sinks (external calls)
23Information flow / data egress
24Key usage / signing
25Rendered screen the human signs
26Autonomous fund movement
99Covert-channel egress
101Observation-to-action atomicity
110Route-class check
122Control-plane write guard
PLANE 6

MULTI-AGENT

9 doors · 27–31, 79, 104, 117–118

Agents acting with agents: hops, coalitions, second principals.

27Agent-to-agent messages
28Delegation chains (multi-hop authority)
29Coalition / collusion
30Orchestrator / planner integrity
31Shared scratchpad / blackboard
79Delegation depth and authority decay
104Inter-agent contention
117Delegation chain
118Second principal
PLANE 7

IDENTITY & AUTHORITY

9 doors · 32–34, 81, 100, 102–103, 119, 127

Who the agent is, and whose authority it carries.

32Agent identity / credentials
33Capability grant / mandate / containment
34Dormant triggers / backdoors
81Agent origin attestation
100Legal-person binding
102Consent grant control
103Delegated-authority persistence
119Agent signer grade
127Token refresh re-attestation
PLANE 8

RUNTIME · INFRA · SUPPLY · TIME

14 doors · 35–42, 85, 89, 92, 108, 113–114

The machine around it: dependencies, triggers, clocks, privileges.

35Scheduler / cron / trigger
36Sandbox / container
37Network egress / proxy
38Telemetry / log pipeline
39Behavioural drift over time
40Dependencies / packages (npm · pip)
41Agent-framework CVEs
42Session handoff / context summarisation
85Inbound trigger authentication and attribution
89Time-source integrity
92Telemetry hygiene
108Kill-switch efficacy
113Capability divergence
114Privilege layering
PLANE 9

TRANSPORT — the wire between components

7 doors · 43–48, 125

The wire between components, and who owns the other end.

43Agent → model request channel
44Model → agent response channel
45Agent → tool/executor call channel
46Tool → agent response in-transit rewrite
47Streaming / token-stream integrity
48Router / gateway trust boundary
125Destination continuity
PLANE 10

BREAKOUT — the Containment Witness

4 doors · 49–52

Declared isolation against observed reach.

49Declared isolation vs independent egress observation
50Persona forgery
51Coalition re-formation after quarantine
52Eval integrity
PLANE 11

CHAINGRAPH — composition, tempo, scope

17 doors · 53–68, 107

Closed at one hop, open at two: the composition.

53Chain composition
54Chain-context ranking
55Choke-point selection
56Machine-speed plan
57Tempo
58Medium
59Autonomic arbitration
60Emergent capability
61Deception density
62Defensive-loop completeness
63Transitive egress
64Rebuild efficacy
65Economic irrationality
66Simulation framing
67Observer invisibility
68Scope verification
107Monitor robustness
PLANE 12

MACHINERY — the runner proves itself

10 doors · 69–74, 76, 82, 93–94

The runner proves it is the runner that was enrolled.

69Orchestrator attestation
70Inference-runtime attestation
71Harness SDK self-attestation
72Template-engine injection
73Unsafe deserialisation
74Runtime module attestation
76Tokenizer + chat-template integrity
82Self-extension
93Signing-key custody
94Entropy source integrity
PLANE 13

SILICON — where the model actually ran

1 door · 75

Where the model actually ran.

75Enclave quote
PLANE 14

HUMAN OVERSIGHT — measured, not assumed

4 doors · 86–88, 106

The human in the loop, measured, not assumed.

86Oversight efficacy
87Approver device binding
88Mandate widening by the agent
106Human-proxy coaching
PLANE 15

ECONOMY — settlement, wallet, contract, meter

8 doors · 77–78, 80, 95, 109, 111–112, 116

Settlement, wallet, contract, meter, rate.

77Agent-to-agent settlement
78Denial-of-wallet
80Autonomous contracting
95Meter reconciliation
109Settlement finality
111Spend-union
112Mandate-bound signing
116Rate ceiling
PLANE 16

LIFECYCLE & LAW — erasure, retirement, the legal person

4 doors · 90–91, 96, 126

Erasure, retirement, re-enrolment, the legal person.

90Lawful erasure vs immutable proof
91Decommission attestation
96Underwriter-verifiable export
126Baseline re-enrolment
PLANE 17

THE UPDATE CHANNEL — every part that changes version

5 doors · 120, 123–124, 128, 131

Every part that changes version: the package, the extension, the tag, the pipeline.

120Version admission
123Extension and plugin update gate
124Immutable reference
128CI context integrity
131The signed config

The word is everyone's. The number is ours.

Every door above is live in the harness. Count the doors your agents pass through, then ask who is checking them.

Map my agents See INTEGRITAS