← All concepts

Paper Compute Concept

Continuous Agent Improvement

A feedback loop: capture relevant sessions, analyze patterns, extract reusable artifacts, apply them to later runs, and measure whether outcomes improve. The model can stay the same. The system around it gets better.

Published May 5, 2026 · Updated May 19, 2026
Improvement Agents Workflow Feedback Loop

Definition

Continuous agent improvement is the feedback loop that uses captured sessions, telemetry, analysis, reusable skills, and review to improve future agent behavior by changing the workflow around the model rather than the model itself.

Continuous agent improvement is not model memory and not fine-tuning. It is the operating loop around the agent.

The loop captures what happened, analyzes patterns, turns useful patterns into reusable artifacts, applies those artifacts to later runs, and measures whether outcomes improve.

The model may stay the same, but the system around it gets better: prompts, skills, tool configuration, evaluation, access patterns, and team knowledge.

Quick breakdown

The five moves
Capture Record relevant agent sessions, including prompts, responses, tool-use events, results, and metadata.
Analyze Identify repeated failures, costly loops, successful patterns, drift, and reusable workflows.
Extract Turn validated patterns into reusable artifacts such as skills, prompts, runbooks, tests, or configuration changes.
Apply Make those artifacts available to future runs through the agent's context, repo, runtime, or workflow configuration.
Measure Track whether future runs improve in success rate, cost, speed, reliability, or reuse.

Why agents don’t improve by default

Without a feedback loop, every session is effectively a cold start. The model doesn’t remember what worked last week, and unless the team captures it, neither does the team.

Models do not automatically retain your organization’s runtime discoveries between sessions. A Claude call today does not know what a teammate’s Claude call learned last Tuesday unless that information is captured and reapplied. Many agent workflows are effectively stateless unless teams add memory, retrieval, skills, or shared context.

Documentation can fill some of the gap, but documentation decays, and many agent sessions produce insights too granular to bother writing up — “the topic needs to be created in the UI first,” “the env var has to be the name, not the value.” These insights tend to matter at runtime and disappear when the session ends. The result is that teams can end up re-solving the same problem repeatedly, with the aggregate cost invisible in any dashboard that does not correlate sessions against each other.

The loop

The improvement loop
   ┌─────────────────────────────────────────────────────┐
 │                                                     │
 ▼                                                     │
Capture ──► Analyze ──► Extract ──► Apply ──► Measure
(captured  (patterns)  (reusable    (future
record)                artifacts)   runs)
 │                                                     ▲
 └─────────────────────────────────────────────────────┘

The loop only closes when all moves are implemented. Capture without analysis produces an unread archive. Analysis without extraction produces insight that dies in a dashboard. Extraction without application produces a library nobody uses. Application without measurement produces vibes, not improvement.

Each move has a concept-level implementation:

  • Capture — a capture layer records the agent-facing sequence and relevant metadata. See telemetry for agents for the capture layer in detail.
  • Analyze — search, queries, dashboards, or reviews identify repeated failures, expensive loops, and successful workflows.
  • Extract — useful patterns become reviewed artifacts: skills, prompts, runbooks, tests, policy rules, or tool configuration.
  • Apply — future runs load or reference those artifacts through the agent’s context, repo, runtime, or workflow.
  • Measure — teams track whether the artifacts reduce repeated failures, tokens per successful task, time to resolution, or human intervention.

Continuous agent improvement is not model retraining

Fine-tuning is sometimes the right answer to “my agent isn’t good enough.” It is not the only one. Continuous agent improvement is a different discipline that updates the system around the model.

Changing the model vs changing the system
Fine-tuningContinuous improvement
Changes model weightsChanges prompts, skills, retrieval, tool configuration, policies, and workflows
Useful for broad behavior shifts or domain adaptationUseful for organization-specific procedures, recurring fixes, and workflow reliability
Requires curated data and evaluationRequires captured sessions, review, reusable artifacts, and measurement
Slower to update and harder to inspectFaster to update and easier to review when artifacts live in version control
Model/provider-specificMore portable when artifacts are written against stable tools and procedures

These are not mutually exclusive. Fine-tuning can improve baseline model behavior; continuous agent improvement improves how a team applies the model to its own systems. Teams should not retrain a model every time they rediscover a workflow mistake. Many recurring fixes belong in skills, prompts, runbooks, tests, or tool configuration first.

What “better” actually measures

“Better” in continuous agent improvement means measurable improvement in task success, cost, speed, reliability, or reuse — not a vague sense that runs feel smoother.

Practical metrics teams track:

  • Resolution rate — what fraction of tasks reach a successful end state.
  • Repeated failure rate — how often the same error signature appears across sessions.
  • Time to first useful tool call — how fast the agent picks a productive path.
  • Time to resolution — total wall-clock from task start to completion.
  • Tokens per resolved task — token spend divided by successful completions.
  • Cost per resolved task — total cost divided by successful completions.
  • Human intervention rate — how often a person had to step in.
  • Skill or artifact invocation rate — what fraction of runs invoke at least one reusable artifact.
  • Regression rate after a skill or workflow update — how often a change makes future runs worse before it makes them better.

Different teams emphasize different metrics. The discipline is to pick a few, baseline them, and check whether the loop moves them.

What continuous agent improvement does not mean

The phrase invites overreach. Specifically:

  • It does not mean the model remembers everything.
  • It does not mean every session should become a skill.
  • It does not mean every run will be better than the last.
  • It does not mean fully automated self-improvement without review.
  • It does not replace evaluation.
  • It does not remove the need for privacy, redaction, access control, and retention policies.

The loop is an operating practice, not a guarantee.

Failure modes

The loop is easy to describe and surprisingly easy to half-implement. Common failure modes:

  • Capture-only. Telemetry turned on, nothing downstream. The result is a searchable archive of problems nobody has extracted.
  • Metric theater. Teams collect dashboards but never change future runs.
  • Overfitting to one session. A one-off workaround becomes a brittle skill that fires in situations it does not actually cover.
  • Extract without review. Generated artifacts go into the library without human inspection.
  • Unreviewed automation. Artifacts are applied to future runs before anyone has validated them.
  • Skills nobody invokes. The library exists; the agent never loads it. Often a trigger-authoring problem.
  • No measurement. Teams extract artifacts but never check whether outcomes improve.
  • Drift. The environment changed. The artifact keeps firing and now produces the wrong fix.
  • Access creep. Captured sessions contain sensitive data but no one governs who can read them.

Each failure mode is fixable, but each requires explicit attention. The loop only improves behavior when someone closes it.

How continuous improvement compounds across a team

The compounding effect of the loop is a team mechanism, not an individual one:

  • Captured sessions create evidence.
  • Evidence reveals recurring patterns.
  • Reviewed patterns become reusable artifacts.
  • Reusable artifacts shape future runs.
  • Future runs produce better evidence.

The longer the loop runs across more engineers, the more the team’s knowledge stops being bottlenecked by any single person’s experience. This is the bridge to team-shared agent knowledge.

How Paper Compute implements continuous agent improvement

Paper Compute implements continuous agent improvement through captured sessions, search and replay-style inspection, and skill generation. The paper CLI uses tapes to capture supported Claude Code sessions routed through the proxy path. Captured records can be searched and inspected; recurring patterns can be turned into versioned skills that teams review and commit so future runs can reuse them. The skill drafting workflow is consolidating under the paper CLI. stereOS provides an isolated runtime for agent workflows that need to execute code or tools safely.

Continuous agent improvement is the operating loop that turns captured agent work into reviewed, reusable knowledge for future runs.

Frequently asked questions

Does continuous agent improvement require a specific framework? +
No. The loop requires capture, analysis, reusable artifacts, review, application, and measurement. Teams can implement those pieces with proxy-layer capture, SDK instrumentation, logs, traces, prompt libraries, skill systems, repos, internal tools, or managed platforms. The important requirement is that future runs can inherit what previous runs taught the team.
Is continuous agent improvement just DevOps for agents? +
It borrows from DevOps: feedback loops, instrumentation, review, automation, and continuous improvement. But the artifact is different. DevOps often improves software delivery and operations; continuous agent improvement improves agent behavior by updating the context, skills, prompts, tools, policies, and workflows the agent uses.
Can the loop be automated? +
Capture, analysis, and draft extraction can be automated. Review should remain human for artifacts that change future agent behavior. Fully automated loops can propagate bad patterns faster than humans can notice them.
Does improvement compound across projects? +
Improvement compounds across projects when reusable artifacts avoid project-specific assumptions or clearly state their boundaries. Skills that encode stable patterns can travel. Skills that depend on local paths, hidden credentials, or one team's environment should stay scoped.
Is continuous agent improvement the same as agent memory? +
No. Agent memory stores information for later recall. Continuous agent improvement is broader: it includes capture, analysis, reusable artifacts, review, application, and measurement. Memory can be one input to the loop, but it is not the loop.
How do teams know the loop is working? +
The loop is working when repeated failures decline, known tasks resolve faster, token or cost per resolved task drops, human intervention decreases, and reviewed skills or artifacts are actually used in future runs.

Where to go next