← All concepts

Paper Compute Concept

Team-Shared Agent Knowledge

One engineer's breakthrough session is a dead artifact until the team can search it, replay it, and invoke the skill extracted from it. Team-shared knowledge is the infrastructure that makes that possible.

Published May 19, 2026
Teams Knowledge Sharing Agents Collaboration

Definition

Team-shared agent knowledge is the infrastructure that turns individual agent session outcomes into searchable, replayable, reusable assets available to every engineer on the team — so one person's solved problem stops being one person's solved problem.

Quick breakdown

What team-shared knowledge requires
Shared capture Every engineer's sessions land in a storage tier the team can query.
Searchable index Semantic and keyword search across every session, scoped by project, author, or error.
Replay Any engineer can step through any recorded session to see how it ran.
Skill library A git-versioned directory of skills extracted from the team's best sessions.
Access controls Project boundaries and provider-key scopes so shared does not mean global.

The knowledge problem in agent teams

One engineer discovers a breakthrough sequence. Without a shared system, that knowledge stays in their head, or in a Slack thread that scrolls out of view.

A software team figured out how to ship code without this problem a long time ago. Git, pull requests, code review, CI. Every change is legible to everyone. An agent team on its own doesn’t inherit any of that infrastructure. Sessions happen on individual laptops. Transcripts land in per-user SQLite files. The patterns one person discovers — the right prompt for a tricky refactor, the exact tool sequence to debug a flaky test, the workaround for a provider’s rate-limit behavior — are locked to that user’s machine unless the team deliberately lifts them out.

The cost isn’t conceptual. It’s wall-clock engineering time. For example, a team of five running ten sessions a day produces fifty daily opportunities to capture a reusable pattern. Most teams capture zero of them. The forty-nine that fall through the floor become the substrate for a slow tax: every month, the team can end up collectively re-solving problems it already solved.

What shared knowledge looks like in practice

The shared layer
Engineer A           Engineer B           Engineer C
  │                    │                    │
  ▼                    ▼                    ▼
Session             Session              Session
  │                    │                    │
  └──────────┬─────────┴─────────┬─────────┘
             ▼                   ▼
      Shared tapes store   Shared skills repo
             │                   │
             ▼                   ▼
      Search / replay      Loaded by next run

Concretely, a team with shared knowledge infrastructure can:

  • Find the session where a teammate fixed the Kafka timeout last sprint, by describing the problem in natural language.
  • Replay a colleague’s debugging run to see which tool calls worked and which failed.
  • Invoke a skill extracted from any teammate’s session, without knowing it existed.
  • See, at a glance, which patterns are invoked most often across the team’s projects.

None of these moves require asking a person to remember anything. The archive is the memory.

Individual vs team-shared agent operation

Two operating modes, same model
Individual modeTeam-shared mode
Sessions stored locally on each laptopSessions replicated to a team-accessible store
Insights communicated in Slack, if at allInsights searchable across every teammate's history
Skills recreated per project by each engineerSkills checked into a shared repo, versioned
Onboarding: "ask someone who's done this before"Onboarding: "search the archive and invoke the skill"
Team capability = strongest individualTeam capability = union of every engineer's best sessions

The second column is not more software. It is the same software, pointed at a different storage tier, with intentional practices layered on top.

How sharing actually scales

For example, a team of 5 engineers each running 10 sessions a day generates 50 daily opportunities to capture a reusable pattern. Without shared infrastructure, 49 are lost.

The compounding effect tends to take a quarter or two to become visible. In the first week, a shared archive can look like a curiosity. By week four, a new engineer can search for “how do we handle Postgres timeouts” and find real examples. By week twelve, many sessions invoke a skill that originated from someone else’s work. The team’s effective velocity stops being bottlenecked by any individual’s experience.

This is the compound interest argument for treating agent sessions as team data. The ROI of the first captured session is tiny. The ROI of the thousandth is the reason a team adopts the practice. The mechanics of that compounding live in continuous agent improvement.

Where sharing has to stop

Team-shared does not mean ungoverned. A shared store that spills credentials, leaks customer data, or mixes unrelated project contexts creates more problems than it solves. Practical boundaries:

  • Project scoping. Sessions tag themselves with a project identifier. Search queries default to the caller’s projects. Cross-project access is explicit.
  • Credential hygiene. Keys used by one engineer’s session do not leak into another engineer’s replay. The proxy redacts at capture time.
  • Retention. The team decides how long sessions persist. Skills persist; raw sessions age out.
  • Sensitive data. Some sessions — security incident response, customer-support replays — require private storage or deletion. The infrastructure supports exclusion lists at capture.

Sharing is a deliberate choice per session, per project, per class of data. The infrastructure makes the default sensible and the exceptions explicit.

Where Paper Compute fits

tapes is the capture and search layer; skill drafting is consolidating under the paper CLI; the shared skills repo is the team library; stereOS is where skills run when they involve executing code in an isolated environment.

Individual mode is the default. Team-shared mode is a practice. The shift from one to the other can be a high-leverage move for a team running agents.

Frequently asked questions

Do you need to run a shared backend for team-shared agent knowledge? +
Not to start. A shared store can be as simple as a git-tracked SQLite file or an S3 bucket the team syncs to. Teams graduate to a managed backend when cross-session search performance or access control requires it. The architecture is local first, shared as a deliberate export.
Is team-shared agent knowledge different from Notion or a shared docs site? +
Yes. A docs site captures what a human wrote after the session. A shared session archive captures what the agent actually did during the session — every prompt, every tool call, every error. Docs explain intent. Sessions preserve behavior. The skill library is the artifact that turns preserved behavior into something the agent can invoke directly, which is what a docs site cannot do.
What about privacy across engineers? +
The same model that protects a developer's local codebase from a cloud LLM applies here. Sessions can be scoped to the author, to a project, or to the org. The proxy redacts at capture time. The store enforces the scope at read time. Shared is a policy decision, not a technical default.
How is this different from LangSmith's team features? +
LangSmith provides shared tracing dashboards. Team-shared agent knowledge goes a layer further: not just everyone can see the traces, but every session can be searched, replayed, and extracted into a skill the next run invokes. The artifact — a versioned skill — does not exist in the LangSmith model.
Can you enforce skills as part of the team's agent runs? +
Yes. Skills are files in a repo. CI can require that every agent run references the skill library. Code review can gate skill additions and updates. The same tooling that governs your codebase governs the skill library — by design.

Where to go next