Can your next agent use what the last one learned?

An agent context layer earns its place when a later agent can retrieve useful knowledge and apply it correctly. A convincing answer is only part of the test. The work still needs to pass a check you defined before the agent began.

This guide gives individual developers and engineering leads a repeatable way to evaluate shared context across the coding agents they already use. It is a test protocol, not a published Firmament benchmark. There are no measured performance results on this page.

Choose one lesson with a checkable consequence

Pick a sandbox task where an earlier discovery matters. A useful candidate is a new feature whose design should reflect an earlier product decision and its reason. Our design decision example shows the distinction: the current interface is visible in code, but the reasoning behind it includes approaches the team already tried and rejected.

Define the review criteria before the run. For a contribution flow, check whether the agent retrieves the earlier tradeoff, applies it within its scope and explains any departure. Fix missing checks or known bugs directly in the product; choose a lesson that remains useful across later tasks.

Write down the starting files, the expected behavior and the command or manual check that decides success. Keep credentials, customer data and production changes out of the experiment. Use a small synthetic example if you cannot share a real task safely.

The lesson should include what worked, why it was necessary, how it was verified and where it applies. “Deployment is tricky” is not enough context for another agent to act on.

Run a baseline before contributing the lesson

Give an agent the task with its normal tools and project instructions. Keep the answer out of its prompt. Save its output, tool calls and check result, including any help you supplied.

The baseline may already pass. The repository, a rules file or the agent's own memory may contain everything needed. That is useful evidence: this task has not demonstrated an additional benefit from shared context.

Let agent A contribute what it learned

After agent A solves the task and verifies the result, ask it to record the durable lesson. A suitable instruction is:

Record the reason this approach worked, the failed alternative, the check that verified it and the project it applies to. Distinguish the lasting procedure from values that must be read live.

In Firmament, an agent contributes through submit. Wait for processing and inspect the resulting wiki page. A successful submission request alone does not prove the lesson is ready for retrieval. Record whether you edited or corrected the resulting page.

Start agent B without the earlier conversation

Use a fresh session, ideally in the other coding tool you actually use. Give it the same starting task and normal instructions to consult project knowledge. Do not paste agent A's transcript, the lesson or the expected fix.

A neutral prompt is:

Work on [task] in [project]. Consult the available project guidance before choosing an approach. Explain which guidance you used, carry out the task and run [check]. If the guidance is missing or conflicting, say so.

Keep the normal instruction to consult Firmament in both trial configurations. In the baseline, the new lesson has not yet been contributed. In the later run, it has. Note any other differences, including model, client, repository changes and local memory.

Ask B to identify the retrieved guidance it used. Then inspect the actual changes and run the check yourself. Naming the right lesson and doing the right thing are separate outcomes.

Record the result, including failures

Check	What to record
Discovery	Did B call the context source without being given the answer?
Retrieval	Did it receive the relevant lesson? Save the response or page reference.
Application	Did its actual work follow the applicable guidance?
Task result	Did the predefined check pass? Keep the output.
Human help	How many corrections or extra instructions were needed?
Time and cost	Include contribution, processing, retrieval and task execution where measurable.
Contamination	Did the answer also appear in local memory, changed files or the prompt?

Report failures in each row. If the lesson was retrieved but ignored, the contribution and retrieval path worked while application failed. If the test passes with no retrieval, the result does not establish that shared context caused the success.

Test an outdated lesson

Change the sandbox so the old procedure is no longer appropriate. Contribute the correction with evidence and inspect the updated page. Start another fresh session and check whether it receives and applies the current guidance.

Firmament processes new submissions into wiki pages that people can inspect and review. Do not assume that a system automatically knows a stored claim has become false. Your evaluation should show how the new evidence reaches it and what happens afterward.

Test team access separately

Two sessions under your own account test reuse across sessions or tools. They do not establish that a teammate has the right access.

For a team test, use a second authorized person and a synthetic lesson in a project intended for sharing. Confirm the project is shared with the appropriate team, then repeat retrieval as that person. Also check that a separate, unshared synthetic lesson is unavailable to them. Avoid using real confidential material as an access-control test fixture.

Read how sharing works and the governance overview before choosing the scope of the test.

What can you conclude?

One successful run shows that this lesson was reused in this workflow. It does not establish a percentage improvement in productivity, token savings or task success across your team.

For a stronger comparison, repeat on representative tasks, keep the scoring fixed and include unsuccessful runs. Separate same-account reuse, cross-tool reuse and sharing between people. Record tool versions and dates so someone else can repeat the experiment.

If your current rules and built-in memory already cover the work, keep using them. Add shared context where your evidence shows a useful lesson failing to reach the next agent. See Firmament alongside rules files and built-in memory.

Try it with your next task

Create a Firmament account, connect your agent and choose one lesson worth reusing. The first useful outcome is simple: your next agent finds it, applies it and passes the check without you explaining it again.