Abstract
The first widely adopted consumer agent may enter our lives through the conversations we already have. Before people trust an agent to represent them to someone else, they need to see that it can understand a situation, retrieve the right context, accept correction, and offer useful help. Human messaging provides a familiar setting in which that trust can develop through small, observable interactions.
This paper proposes a multi-messenger as the foundation for that experience. Relationships continue across applications, while their records remain fragmented. A personal agent needs a way to connect authorized context around people and situations, without requiring users to continually explain their lives. The product should be useful immediately as a messenger, then earn a larger role through remembering, interpreting, drafting, and carefully bounded action.
The argument grows out of my work on Collision Chambers and Collisions. It is a product and architecture thesis, not a claim of demonstrated market adoption. The central proposition is that relationship continuity can make assistance useful, and visible usefulness can create the conditions for delegation.
1. Before two agents negotiate, one must earn trust
Earlier in 2026, I worked on Collision Chambers, an experimental approach to agentic negotiation. The question was how agents representing different people might explore shared goals and tensions before either person disclosed sensitive context. I wanted to preserve each person’s stakes while making common ground easier to discover.
I struggled to explain the product. At first, I thought the language was the problem. With distance, I began to suspect that the sequence was wrong: I was explaining what agents could do together before establishing why someone would trust even one of them.
Agent-to-agent negotiation asks a person to accept several unfamiliar propositions at once. My agent understands me. It can represent my intentions. Your agent can represent yours. Their interaction will preserve what matters to both of us. And delegating that interaction will be more useful than simply talking.
In the situations I considered, the strongest reasons to accept those conditions often involved difficult, infrequent, high-stakes conversations—the very situations in which mistakes would be hardest to tolerate. A compelling coordination mechanism still needs a credible path to everyday trust. That led me back to the human conversation already happening before any agent arrives.
2. Human conversation is the starting interface
People already use messaging to coordinate, ask for help, apologize, flirt, persuade, make plans, and change their minds. The interface is familiar, and the motivation comes from an existing relationship. An agent can help with a concrete difficulty inside that activity: a forgotten commitment, a confusing exchange, or a reply someone cannot quite phrase.
By an agent inside human chats, I mean private assistance alongside a conversation. The agent does not need to become a visible participant in the group or speak to the other person. It can help the user recover context and consider a response while leaving the human conversation intact.
The first useful interaction might be small: “What did we agree on last week?” or “Help me explain why I need more time.” The user can inspect the answer against a situation they know. That makes the agent’s understanding—or misunderstanding—visible.
This is a different starting point from an onboarding flow that asks someone to describe their personality, upload their history, grant broad access, and wait for future value. A bounded task gives the person a reason to share bounded context. The agent earns another interaction by helping with this one.
3. People already produce the material
I am interested in the idea that everyone is already an author of their own life. Most people simply do not turn that life into a polished artifact. Their observations, preferences, questions, jokes, and decisions remain in unfinished notes and scattered conversations.
Making oneself legible usually takes work: selecting what matters, imposing a sequence, explaining the background, and making contradictions understandable. I do not want to maintain a perfect second brain before a personal agent can help me. I want to speak naturally, leave thoughts unfinished, make typos, and still be understood well enough for the task at hand.
A message to a friend and an exchange with an aunt may reveal different parts of the same person. Sometimes those contexts are relevant to each other; often they are not. The opportunity is to recover useful connections without asking the user to produce a second account of everything that happened.
Conversation is rich source material, but it is not a complete or neutral autobiography. People perform different roles, joke, omit things, contradict themselves, and evolve. An agent should treat the record as partial evidence, not as a definitive model of someone’s identity. A person must be able to say, “That was true then,” “I was joking,” or “You have misunderstood what mattered.”
4. Why a multi-messenger?
A relationship does not restart when it moves between applications. A conversation may begin on Instagram, continue on WhatsApp, and turn into a plan over email. The people retain continuity, but each application stores only its own portion. Remembering how those pieces connect becomes the user’s job.
An assistant confined to one application inherits that boundary. It may answer accurately from the records it sees while missing a change of plan or a qualification expressed elsewhere. Bringing more data into a prompt does not solve this by itself. The system also has to know which conversations belong together, where each statement came from, and what remains missing.
A multi-messenger creates a practical place to do that work. It gives the user an immediate reason to return: talking to people they already know. Authorized conversations can then provide context for assistance inside those same relationships. The user does not have to develop a separate habit of documenting life for an assistant.
This creates a product loop worth testing: existing relationships bring people to the messenger; relevant context improves assistance; useful assistance makes the messenger worth returning to. Each part has to work. An unreliable inbox will lose the activity that provides context. An intrusive agent can make an otherwise useful messenger less desirable.
A multi-messenger is not the only possible implementation. Extensions, operating-system integrations, and selective imports could also connect context. My reason for building a messenger is the proximity between context and action: the user can read the conversation, consult the agent, and return to reply in one place. That benefit must justify the cost of connecting accounts and adopting another client.
5. Understanding must become visible before delegation
Reassurance alone cannot establish trust. A person needs evidence that the agent understood the relevant situation and will remain within the authority it was given. I propose a progression in which each step can be useful on its own:
- Remember. Retrieve a relevant message or commitment and show where it came from.
- Interpret. Offer a tentative account of the situation, with room for correction.
- Advise. Suggest possible responses and explain the tradeoffs in relation to the user’s stated goal.
- Anticipate. Surface something that may matter next, as a suggestion rather than a claim to know what the user wants.
- Assist. Prepare a draft or action preview that the user can edit.
- Act. Carry out a specifically authorized action, with a clear recipient, scope, and record of what happened.
This is a progression of possible responsibilities, not an automatic permission ladder. Someone may want excellent recall and drafting forever, without allowing the agent to send anything. Accepting a suggestion does not authorize the next action.
The feeling of being understood can be powerful. It can also exceed what the system has actually demonstrated. Warmth and fluency are not evidence of accuracy. Trust should grow from correct retrieval, useful interpretation, honest uncertainty, and successful correction—not simply from a convincing conversational style.
6. A concrete implementation in Collisions
Collisions is my working implementation of this direction. I have worked on its messaging client and personal-agent architecture. The current client connects through Beeper; a local gateway connects the interface to the agent runtime. The implementation separates external messages, local organization, and agent conversations.
Keep the original conversation authoritative
The inbox brings connected chats into one interface. Chats associated with the same person can appear in a combined timeline, while the user can still select an individual source. Replies go through the selected messaging service. Original records retain their source identity; the combined view does not turn them into a new, independent conversation.
The client uses a local message mirror and retrieves older history progressively. That makes available context a matter of coverage, not an assumption that the agent has read everything. A useful answer should distinguish what it found from what it has not yet retrieved.
Organize context around people and Spaces
Spaces bring together people, notes, saved sources, and agent sessions. They provide continuity across individual exchanges without requiring every past message to be inserted into every prompt. The personal agent can retrieve relevant material when a task requires it.
The memory architecture retains references to source evidence and distinguishes explicit information from inferred observations. Corrections and deletions in a source invalidate affected derived memory. The goal is a revisable account of the situation, rather than an accumulating collection of claims that becomes harder to correct over time.
Separate assistance from external speech
The interface can open the personal agent beside a human conversation. A user can bring relevant messages into that private exchange, ask for help, and work on a response before returning to the messenger. Agent history is stored separately from the original chat. Provider execution sessions are separate from the application’s saved history, so continuity is owned by the application rather than by a transient model session.
The visible outcome is a workspace for reading, thinking, and replying. The current prototype demonstrates that interaction structure. It does not establish that users will prefer it, that inferred memory is always correct, or that agent access is automatically limited by the Space visible on screen. Spaces organize context; enforceable access boundaries require explicit controls.
7. Existing context is not automatic permission
The fact that people already generate intimate material does not entitle an agent to use it. A message was written for an audience and a purpose. Reading it through an assistant, retaining an interpretation, or sharing a conclusion can change both.
Onboarding should therefore connect a concrete benefit to a comprehensible scope: which accounts or conversations will be used, what will be retained, and what the agent can do with that context. People should be able to begin with a small selection and still receive useful help.
Other participants matter too. A user’s access to a conversation is not unlimited authority to expose another person’s private information. Permission to retrieve context for a private draft should not quietly become permission to disclose that context to another person or their agent.
More context is valuable only when it is relevant, interpretable, and appropriately authorized. A mistaken connection between identities, an old preference treated as current, or a joke treated as a stable intention can make the agent more confident and less helpful. Source visibility, correction, retention choices, and clear action previews are part of the product’s usefulness.
8. What would make the thesis credible?
The decisive question is whether relationship continuity helps people do something they already care about. Message volume, the size of a memory store, and fluent responses cannot answer that question alone. I would evaluate the thesis through a few concrete comparisons:
- Recall: Can people recover a relevant agreement faster than by searching their existing apps, and is the cited evidence correct?
- Context: Does authorized cross-application history improve an answer compared with the active thread alone? Does irrelevant history make it worse?
- Correction: When a person rejects an interpretation or changes a source, does that correction affect later assistance?
- Drafting: Does the agent help users express what they intended, with edits they consider worthwhile, while leaving them in control of the final message?
- Return use: Do people return for relationship tasks after the novelty fades, and is the benefit worth the setup and trust cost?
- Bounded action: Does the system reliably stay within the recipient, content, and action the user approved?
These are proposed evaluation criteria, not reported results. The thesis would weaken if selective context sharing worked just as well with much less friction, if cross-application memory mostly added noise, or if users found a separate assistant preferable to changing their messenger. A useful experiment must leave room for those outcomes.
9. Coordination can follow earned understanding
I still see a place for agent-to-agent coordination. The order matters. A person should first have an agent whose understanding they can inspect and correct, then a specific reason to let it coordinate with someone else. The scope of that exchange needs to be chosen, rather than inferred from the amount of context the agent possesses.
Human conversations are where intentions are already expressed, tested, negotiated, and changed. They offer a grounded place to learn what assistance people want and what representation they will trust. An agent that can help someone recover a forgotten commitment has done something observable. An agent that can help them articulate a difficult response has earned another kind of usefulness. Neither achievement gives it an unrestricted mandate.
I am building toward a personal agent that makes the life someone already lives easier to understand and act within. A multi-messenger gives that work a concrete setting: existing people, ongoing conversations, and decisions that matter now. The agent earns its place by helping with those decisions, one understandable step at a time.