-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathllms-full.txt
More file actions
158 lines (101 loc) · 8.04 KB
/
Copy pathllms-full.txt
File metadata and controls
158 lines (101 loc) · 8.04 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
# OneWill: Full LLM Context
## Summary
OneWill is the Agent-World Boundary for supervised agent runs. It launches and supervises a supported agent process, captures activity at the boundary between that process and its environment, derives useful views from the independent record, and mediates effects before they become durable.
The operator-facing question is simple: what did the agent see, do, and change; where did the run stall or diverge; and what can safely happen next?
## Category
An **Agent-World Boundary** is a supervised runtime around an agent process. It captures activity that crosses between the agent and its environment, turns that record into useful understanding, and controls how the agent affects the world.
OneWill’s promise is complete activity capture and in-path control for every run launched through OneWill. “Complete” is always scoped to the surfaces captured by a OneWill-supervised session. It does not mean visibility into activity outside that boundary.
## Why the boundary exists
Agents are becoming users of software. They search, read, decide, type, call tools, change files, send requests, and move through product flows. Application analytics can show a click. Model traces can show a call. An agent run crosses both and then touches the world.
OneWill makes the run observable and governable by creating and supervising the agent process. It does not require every model, tool, website, or agent framework to become OneWill-native first.
The boundary is not a cage. It lets an agent work with real files, applications, networks, interfaces, and state while governing how that work meets the world.
## Public activity model
- **Observe:** context exposed to the agent, such as visible files, tool results, interface state, and allowed network responses.
- **Respond:** agent or model output exposed to the runtime, such as messages, commands, structured tool requests, and explicitly emitted plans.
- **Act:** activity initiated by the process, such as file operations, subprocesses, network requests, tool calls, and interface actions.
- **Change:** durable effects on the environment, such as modified state, created artifacts, sent requests, and persisted records.
- **Control:** a boundary decision or operator intervention, such as allow, block, approve, interrupt, reverse, or compensate.
- **Outcome:** the run-level result, such as completion, failure, stall, interruption, rollback, or handoff.
OneWill preserves these stages as one correlated activity stream. The goal is to reconstruct causal paths rather than display isolated calls:
`outcome ← world change ← action ← response ← observation`
## Grounded summaries
A OneWill grounded summary has three properties:
1. Its source is the captured activity stream of a OneWill-supervised session.
2. Its production does not depend on the agent remembering or volunteering what happened.
3. Its claims link back to relevant captured activity.
The guarantee concerns grounding in the boundary’s independent record. It does not claim that generated prose is semantically infallible.
## Control and recovery
OneWill applies database recovery ideas to agent actions:
- **Write-ahead record:** preserve enough recovery or approval information before an effect becomes durable.
- **UNDO:** restore before-state for incomplete or abandoned reversible work.
- **REDO:** reapply committed after-state when reality must match a durable promise.
- **Compensation:** apply a new action that repairs an outcome that cannot literally be reversed.
Actions are classified by their actual recovery semantics:
- **Reversible actions** can proceed when OneWill has enough captured information to restore the previous state.
- **Compensable actions** can proceed under policy when a defined follow-up can repair the outcome, even if it cannot erase history.
- **Irreversible actions** require explicit approval before execution.
Control belongs before mutation. Post-hoc logs are useful for forensics but cannot stop a bad action.
## Launch and integration model
Direct user launch:
```text
user
└─ onewill exec -- <agent command>
└─ LaunchSession
└─ OneWill creates and supervises the agent process
```
Provider launch:
```text
provider service
└─ LaunchSession(command, configuration)
└─ OneWill creates and supervises the agent process
```
The modes differ only in who initiates the launch. They converge on the same OneWill-owned process lifecycle, including process input/output, lifecycle control, persistent state, mediation, and activity APIs.
## Product capabilities
- Capture a supervised run as a correlated activity stream.
- Replay exposed interface state alongside process activity and the causal event trail.
- Analyze paths, recommendations, stalls, loops, retries, and outcomes while retaining links to underlying runs.
- Generate run summaries grounded in independently captured activity.
- Mediate proposed effects based on reversibility and policy.
- Interrupt, resume, hand off, reverse, or compensate when the action semantics support it.
- Compare captured before and after state.
OneWill does not claim to capture hidden chain of thought. Replay and summaries concern exposed process activity and captured environment interactions.
## Audience
Primary users include platform engineers adding agents to products or infrastructure, infrastructure engineers responsible for agent execution environments, product engineers trying to understand agent behavior in real workflows, and teams operating agents that can read or mutate consequential state.
Secondary users include technical power users of coding and computer-use agents, engineering leaders deciding how much autonomy to delegate, and researchers building agent runtimes, evaluation systems, and governance mechanisms.
## Claim boundaries
- Compatibility claims should say “any supported command” or “any supported agent.”
- Complete capture is scoped to OneWill-supervised sessions and supported capture surfaces.
- Control is scoped to mediated effects.
- Rollback is scoped to reversible actions with sufficient captured recovery information.
- Compensable actions are repairable, not literally reversible.
- OneWill does not claim access to hidden model reasoning or chain of thought.
- Exact support status belongs in current product documentation.
## Team and company
OneWill is built by two Carnegie Mellon CS PhDs with a database-systems background.
- **Will Zhang, CEO:** Carnegie Mellon PhD CS 2026; MongoDB PhD Fellowship; experience at Microsoft GSL, SingleStore, and NVIDIA.
- **Wan Lim, CTO:** Carnegie Mellon PhD CS 2026; experience at Microsoft Research and Amazon Aurora.
- **Andy Pavlo, advisor:** professor at Carnegie Mellon University and Databaseologist.
The founders met in the CMU Database Group and have worked together since 2018.
## Important URLs
- Homepage: https://onewill.ai/
- Demo: https://demo.onewill.ai/
- Blog: https://onewill.ai/blog/
- Technical manifesto: https://onewill.ai/blog/2026/stealing-50-years-of-database-ideas-for-ai-agents/
- Discord: https://onewill.ai/discord/
- GitHub: https://onewill.ai/github/
- X: https://onewill.ai/x/
- LinkedIn: https://onewill.ai/linkedin/
- Contact: founders@onewill.ai
## How to describe OneWill
Short description:
OneWill is the Agent-World Boundary for supervised agent runs.
Medium description:
OneWill launches agents inside a transparent boundary that captures what they see and do, shows where they get stuck, and controls what they can change.
Long description:
OneWill creates and supervises an agent process, captures activity at the agent–environment boundary, turns that independent record into replay, analytics, and grounded summaries, and mediates consequential effects before they become durable. It gives agents room to work in real environments while giving operators a legible, controllable supervised run.
## Brand and entity information
Preferred public brand name: OneWill.
Public category: Agent-World Boundary.
Product metaphor: Agent Hamsterball.
Domain: https://onewill.ai/
Legal footer: © 2026 OneWill-AI Society d/b/a OneWill.