An experienced colleague rarely starts a proposal with nothing. They know where the current price list lives, which template to use and who needs to review it.
An agent needs that context, too. A capable model can produce the wrong kind of work when requirements remain implicit.
Our starting point for Nodejam is to bring files, tools and instructions into the conversation, then make the result easy to inspect and revise.
Bring the source of truth
A task might depend on quotations, a project record, a calendar entry and an email. These sources have different permissions and levels of freshness.
Connecting a tool does not settle which record is authoritative. The task still needs to distinguish an approved budget from a draft, or the latest brief from its predecessor.
“One place” means a place to work with that context through chat. It does not require every service to disappear or every record to move into a new database.
Show what good looks like
A template can explain more than a long prompt. It may define a report’s sections, a price list’s columns or the information needed on each slide of a sales deck.
An example also communicates house style. How much text fits comfortably? How are images credited? Which details accompany a product recommendation?
Examples need boundaries. A previous document can contain outdated facts. A visual reference may express a preference rather than an industry standard. The agent needs to know what to preserve and what to reconsider.
Turn requirements into review criteria
Some requirements become direct checks. Required sections must exist. Totals must reconcile. References must point to supplied sources. Exported files must open.
Others need judgment. A technically complete sales deck can feel visually incoherent. A proposal can follow a template but miss the client’s concern. These need explicit criteria and an informed reviewer.
A citation format helps check consistency; it cannot establish that a claim is true. A checklist may support a compliance process without proving legal compliance. Each check has limits.
Build on the harness, focus on the work
The coding-agent ecosystem has developed foundations for tool use, file access and iterative execution. We are building on those foundations to focus on what knowledge workers need around them.
That includes a straightforward desktop experience, useful file workflows, connected tools and ways to encode how good work gets done. For a company, it can also include shared coordination, access rules and agreed review paths.
Community contributions might include reproducible file tasks, templates with clear requirements or examples of revisions that broke previously correct results. Contributions need not all change the core harness.
Measure work people actually do
A useful evaluation package includes permitted sources, a brief, expected checks and examples of mistakes that matter. Model-based review needs calibration against informed human judgment, as Anthropic’s agent evaluation guide explains.
This is our philosophy behind “the harness for knowledge work.” Give the agent a useful working environment, make expectations explicit, and help the person receiving the result see what they can rely on.