The Project Agent Harness

An AI agent harness built for project management

Coding agents popularized the harness. TensorPM brings that architecture to project management: a model loop wired to a persistent, typed project graph — goals, scope, risks, budget, timeline, milestones, work packages, action items and decisions — with Context Distillation, governed tools and explicit permission boundaries. Distilled inputs and core project-context proposals require human confirmation; authorized typed MCP CRUD can apply directly. For people that shows up as an AI project agent. For external agents it shows up as MCP context and tools plus an A2A project agent: your harness keeps the loop, ours keeps the project.

PM-specific agent harnessProject graph, not a file treeMCP server + clientA2A project agentACP commerceAppend-only decisionsAgent self-schedulingExplicit approval boundaries2M Trial creditsBusiness BYOK

Why project management needs its own harness

A harness is what turns a model into an agent: the state it works on, the tools it may call, and the rules it runs under. Many coding-oriented harnesses center their work on a repository and file tree. A long-running project needs a different state model — and each of these six elements is a load-bearing part of TensorPM.

01

The state layer is a project graph

Not a folder, not a chat log. Goals, scope, stakeholders, budget, timeline, milestones, work packages, action items, risks and decisions — typed, related, versioned. Every tool operates on that shape, so the agent reasons about the project rather than about text that happens to mention it.

02

Context Distillation instead of prompt stuffing

Mails, messages, meeting notes and documents are distilled into candidate Action Items, Decisions and Risks. Nothing enters the graph before a human confirms it. That is how the context stays complete, current and confirmed — the precondition for every analysis on top of it.

03

Human-in-the-loop by construction

Approval boundaries are explicit. Distilled inputs and core project-context changes use proposals that a human confirms. Authorized typed MCP operations such as action-item CRUD can apply directly and remain visible in project history. The harness states which path applies instead of relying on a prompt to imply it.

04

Decisions are a first-class primitive

record, supersede, withdraw — with source, actor, rationale and timestamp, chained so the audit trail survives every change of mind. Project management is decision management. A harness for it needs a decision primitive, not a comment field.

05

Time is a primitive, too

Milestones, deadlines, check-ins — and agent self-scheduling: the harness can put itself back on the clock and fires with the full project context in hand. A task-oriented agent run may end with the turn; a project does not.

06

Open at both ends

Bidirectional MCP (server and client), A2A, native mail and messenger connectors, and delegation to coding agents. TensorPM does not replace your agent's harness — it adds a persistent project layer that the agent can use.

One system, three precise roles

The terms describe different layers of the same architecture. Their separation makes TensorPM's role clear for people, inside its own runtime, and for external agents.

For people

AI project agent

TensorPM understands the confirmed project context, analyzes risks and dependencies, evaluates options, and provides guidance for the next steps.

Inside TensorPM

Project agent harness

The internal execution layer connects the model loop, persistent project graph, Context Distillation, governed tools, permissions, and human approval.

For external agents

MCP context + A2A agent

MCP supplies governed project context and tools without replacing the client agent's harness. Through A2A, TensorPM participates as a specialized project agent.

Project-agent capabilities

Beyond CRUD on action items, TensorPM provides governed decisions, self-scheduling, delegation, and connectors on top of the shared project graph.

Decisions & commitments

Append-only records with full audit trail

MCP — 6 tools
  • record_decision — capture a stakeholder commit, top-down change, or agent recommendation as a first-class record (source, actor, rationale, decidedAt).
  • supersede_decision — atomically replace an active decision; the previous record is chained as 'superseded' so the audit trail stays intact.
  • list_decisions — query by status (active / superseded / withdrawn), source, or linked entity (action_item / risk / milestone).
  • link_decision / unlink_decision — connect a decision to the action items, risks, or milestones it governs.
  • withdraw_decision — mark a decision withdrawn when no replacement applies. Prefer supersession when there is one.

Agent self-scheduling

The TensorPM project agent can schedule its own future runs

A2A — message/send
  • Send an A2A message/send describing the future intent (e.g. "remind me to review the budget next Tuesday").
  • Useful for follow-ups, milestone check-ins, recurring reviews, and time-bound risk re-evaluations.
  • There is no direct MCP tool for self-scheduling — scheduling lives with the project agent so it has full context when it fires.

Bidirectional MCP, delegation & connectors

MCP server + client, A2A, and native messenger connectors

MCP + A2A + connectors
  • TensorPM is bidirectional MCP (server and client) plus A2A — connect external MCP servers into the project and expose the project graph to any MCP/A2A agent.
  • Delegate tasks to coding agents: GitHub Copilot (live sub-agent call), Codex and Claude Code (via agent assignment of action items, in developer / architect / reviewer roles).
  • Native Email (mail ingest + mail agent) and Telegram messenger connectors let project participants interact with the agent under a per-role permission model (post status, complete action items, propose decisions, or read-only).
  • Incoming messages and mails are distilled into Action Items, Decisions, or Risks and only mutate the project graph after human confirmation (human-in-the-loop Distiller).

One-click install

Add to Claude DesktopOpens the MCP install dialog in claude.aiAdd to CursorOpens Cursor's MCP install flow
macOS (Homebrew): brew install --cask neo552/tensorpm/tensorpm
Windows (winget): winget install --id Neo552.TensorPM --exact --accept-package-agreements --accept-source-agreements
Linux AppImage: curl -fL -o ~/TensorPM.AppImage https://tensorpm.com/api/download/linux && chmod +x ~/TensorPM.AppImage

Agent-friendly pricing

Start on Trial with Cloud Sync already included, move to Pro for more hosted AI credits and unlimited own cloud workspaces, or use Business when API-key/BYOK management and custom limits are required.

Trial · €0

No time limit, 2,000,000 lifetime TensorPM AI credits, Cloud Sync included, several own cloud workspaces under a fair-use limit, unlimited free members, no BYOK/API-key management.

Pro monthly · €99/month

Unlimited own cloud workspaces, 1 active project, unlimited archived projects, and 10,000,000 hosted AI credits per month. Cloud Sync and free members are on every tier. No BYOK.

Pro yearly · €990/year

Same Pro entitlements with yearly billing. 12 months for the price of 10 monthly payments.

Business · custom

Custom limits and credits, Cloud Sync, and the only plan with BYOK/API-key management. Contact sales — no self-service checkout.

Agent-readable docs

Discover & integrate programmatically

Every endpoint and capability is described in machine-readable form. Point your agent at these URLs and it will know what TensorPM can do, how it costs, and how to check out.