Jason Harness gives AI sales agents persistent campaign memory
Reply.io’s MIT-licensed Jason Harness gives AI sales agents nine roles, local campaign memory and human send approvals, while external models and sending services still receive data.

Why it matters
- Reply.io opened Jason Harness’s complete runtime source on November 1 and announced the alpha publicly on November 17.
- The runtime and extension system are available under the MIT license.
- The nine roles describe host-agent jobs and do not necessarily mean that nine separate models run.
- Reply stores campaigns, contacts and history locally, while connected model and sending services may still receive required data.
- The launch provides no independently validated results for deliverability, qualified replies or meetings.
By Rebecca Stone
Reply.io opened the complete runtime source for Jason Harness on November 1, then announced the MIT-licensed sales-agent alpha publicly on November 17. The release gives existing agents a persistent workspace for prospecting, follow-up planning and message drafting, but supplies neither a new model nor independently validated sales results.
Reply.io says the alpha still relies on external models and sending services. That combination defines the market trade: persistent workflow and local campaign records, against provider dependencies that local storage does not remove. For sales teams, the stakes are practical: retain context across chats without assuming a local database prevents data from reaching outside services.
What does Jason Harness add to an existing agent?
Jason Harness does not ship with its own model. It packages instructions, campaign records and approval steps so an existing agent can resume a campaign across separate sessions. Reply.io calls it “an open-source tool that turns general-purpose AI agents into a workspace for sales outreach.”
The alpha can give tools such as Claude Code or OpenAI Codex distinct roles for researching prospects, planning follow-ups and drafting messages. The roles are jobs for a host agent, not evidence that nine separate models run at once. That distinction keeps the open-source harness separate from Reply’s managed Jason AI sales development representative.
The launch describes two separate products rather than a replacement for the managed service. Reply’s launch material supports a narrow claim: the project adds workflow structure around an agent users already have. It does not establish a new model, a new sender or a measured sales improvement.
For teams assembling their own toolchain, Jason Harness functions as a shared campaign layer. It preserves instructions and records across sessions while leaving model choice and delivery infrastructure outside the package. The open-source release also makes both the runtime and its extensions available for inspection and adaptation.
Do nine roles mean nine AI models?
Reply identifies nine roles for the campaign:
- Campaign management
- Planning
- Research
- Writing
- Personalization
- Critique
- Response handling
- Analytics
- Deliverability
These labels define distinct jobs for the host agent. A planner can propose a sequence, a researcher can gather context, and a critic can flag an unsupported claim before a draft goes to a human reviewer. The critic therefore operates inside the drafting flow rather than after a message has been sent.
The count describes task separation, not the number of models running. It also does not provide performance evidence for each function. A host agent can move through several roles because the harness assigns responsibilities within one campaign, rather than establishing nine independent model deployments.
Why keep campaign state outside the chat?
A seller who starts a fresh conversation still needs to know which pitch a prospect saw, what the prospect said in response and whether a colleague already approved a follow-up. General-purpose assistants do not automatically carry that operational history into every new chat.
Reply stores campaigns, contacts and history in a local database, with rules and memory in local files. Like a campaign notebook kept beside the sales team, the record gives a later session continuity without requiring chat history to carry every detail. The package combines instructions, campaign records and approval steps so an existing agent can pick up the work across sessions.
The harness keeps that campaign state on the operator’s computer while leaving delivery to an external sending service. Code for the runtime and extension system is available under the MIT license, according to Reply’s open-source documentation. Developers can inspect and adapt both parts rather than treating campaign memory as a closed hosted feature.
That continuity depends on the campaign record remaining coherent as work moves between roles. A planner’s sequence, a researcher’s findings and a writer’s draft can remain tied to the same stored context. Each new session starts from the campaign files rather than an isolated prompt.
Does local storage guarantee privacy?
Local storage does not guarantee privacy. The underlying agent processes requests according to its own provider and plan, so Jason Harness does not make the tool an offline AI system. Connected services may still receive whatever data they need to complete a task.
Delivery takes place through Reply or another sending provider connected through an extension. The connected provider handles the sending path, which can carry data beyond the local database. Reply’s launch material therefore does not present local storage as a blanket privacy guarantee.
Any organization evaluating the alpha would need to trace what goes to each service. That operational requirement follows from the separation between locally stored campaign state and externally executed tasks. The local database records the state; it does not control every downstream transfer.
For teams evaluating prospect records and campaign histories, that service-by-service trace matters more than the label “local.” Reply has also left sending-provider selection open, allowing operators to use Reply or another provider through an extension. The choice affects both the data path and the operating cost of a campaign.
What does human approval change?
Reply says “a person must approve messages before they go out.” The approval gate can prevent an unreviewed draft from reaching a prospect, but it does not establish that every claim is accurate or appropriate for the campaign.
A message can sound specific while relying on an outdated job title or an unsupported claim about the recipient’s company. The reviewer also needs to know whether someone opted out, whether the proposed follow-up repeats an earlier touch and whether the send still fits the original campaign brief.
Human approval only helps when reviewers can see opt-outs and prior touches alongside each draft. A generic approval control does not resolve missing context. It becomes useful when the draft and the campaign record appear together at the point of review.
Moving drafting and research into the same workspace may reduce context switching. It does not remove the cost of reviewing what the AI produces. The critic role can flag an unsupported claim before the draft reaches the reviewer, but the person approving the send still carries the final judgment.
Is open-source outreach free?
The MIT license makes the runtime and extension system inspectable and adaptable. It does not make outreach free. AI subscriptions may have usage limits, while sending accounts or prospect data can carry separate costs.
Teams choose whether sending runs through Reply or another provider connected through an extension. That choice affects both operating cost and the route taken by campaign data. Those limits remain separate from the software license.
Reply’s launch material does not provide independently validated results for deliverability, qualified replies or meetings compared with a conventional workflow. For prospective users, code access establishes inspectability, not commercial performance. The alpha’s business case therefore rests on measured workflow gains rather than a demonstrated improvement in those sales outcomes.
What should teams measure?
The clearest test for prospective users is whether Jason Harness can reduce the time spent reconstructing campaign context without increasing the time needed to correct it. Teams comparing the tool with an existing workflow could measure:
- Review time per message
- How often drafts are rejected for inaccurate claims
- Whether replies are passed back to the right person
Review time measures the labor surrounding AI output, while rejection rates test factual reliability. Reply routing checks whether campaign state survives an actual prospect response. These measures would say more about the harness’s value than the number of roles it can assign.
Any broader rollout should be judged by those measured changes, because Reply’s current launch provides no independently validated deliverability, qualified-reply or meeting advantage.
Source: unite.ai
More from Rebecca Stone
Show full bio
Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
802 articles
Related articles
- Two Engineers Built Stack Overflow for Agents. ChatGPT Tripled Its Signups
- OpenAI ships a model-native harness and native sandboxes for its Agents SDK
- Mirakl Bets on AI Agents and ChatGPT Enterprise to Rebuild Commerce
- Brian Chesky: AI Agents Need Their Own Operating System
- OpenAI Launches Workspace Agents in ChatGPT for Teams