What TeamShared should be
Teams are starting to run on agents, and the agents don't share anything. Each one has its own context window, its own view of the customer, its own idea of what was decided last week. The people on the team have the same problem, just slower. TeamShared is the fix for that: one memory for a team and its agents, built so that work can run for weeks, across many hands, without losing the thread.
Here's the plan, in four parts.
1. One shared memory
Everything a team learns goes into one place that every agent and every person reads from. Facts, decisions, preferences, how-to skills, playbooks, files. Memory is written through MCP from any harness, recalled by meaning, and distilled from conversations automatically. Two agents editing the same playbook at once get a conflict, not a silent overwrite.
This is shipped. It's the foundation everything else stands on.
2. An ontology of the business
Memory about "the customer" is only useful if everyone means the same customer. So TeamShared keeps a typed registry of what the business is made of: people, systems, vendors, customers, projects, and whatever kinds a team defines. Links between kinds are declared. When two records turn out to be one thing, merging rewrites every reference and keeps the old name as an alias. The whole graph is built from an append-only log, so you can ask what we knew about a system on a given Tuesday.
Changes to the ontology can be routed through review. Automation levels decide how much an agent can change on its own and only a person can raise them.
This is shipped. Each entity has a hub page that gathers everything attached to it.
3. Processes that aim at goals
Long work, the kind that takes weeks and crosses many hands, should not be a flowchart. It should be a goal, a subject in the ontology, a set of constraints, and a starting plan that is allowed to change.
- The goal says what done means.
- The subject ties the work to a real entity, so an agent picking it up on day 19 inherits everything the team knows about it.
- Constraints are the part that doesn't bend: evidence that must exist before a step closes, decisions only a person can make, spend limits, deadlines.
- The starting plan is a playbook the team trusts, used as a default rather than a script.
Agents move the work forward by asking for a context pack, drafting the next action, attaching evidence, fanning out to parallel sub-work, retrying on timers and escalating to a person when they run out of attempts. Every step is an event on an append-only stream. An admin can override a gate with a reason, and the reason is on the record. An agent can't.
Most of this is shipped: instances on ontology entities, context packs, evidence gates with recorded overrides, deadline timers, retry and escalation, fan-out and join, cost rollup across the tree, and human approval for spend, publish and destructive actions.
Not yet: goals as a first-class field with progress against them, structured recording of departures from the plan, inbound triggers from webhooks and email, and sharing templates between orgs.
4. A system that learns without drifting
When work closes, what worked should feed back into the playbooks the next run starts from. The distillation and curation workers already do this for conversations and facts. Pointing them at closed processes, and having a person review each proposed playbook change before it lands, is the next step.
The rule across all four parts is the same: change freely, record everything, keep it reversible.
Why this shape
I learned the standard from the people I worked for who ran security the hard way. Don't trust a control you can't evidence. Treat the record as the deliverable. Build a team that gets calmer after each incident instead of more tired. What they did by hand for one company, every team's agents now need as infrastructure. The ideas here are more theirs than mine.
If you've built teams that work this way, you already know whether this plan is right. Tell me the constraint you'd never let an agent bypass. That's the conversation I want to have.