Back to Research
Technical Report 13 — preliminary, not peer reviewed

Bounded Self-Improvement in an Autonomous Agent Loop

Akash Deep · Aethelas Private Limited, Patna, India
Measuring where autonomous iteration improves, plateaus, and fails
Abstract

Autonomous, self-improving AI is often described as though improvement were unconditional. We built an autonomous loop that chains our system's existing layers — memory, fenced reasoning, planning, tool use, reflection, and a learning controller — into a continuous goal-pursuing cycle, and measured whether each cycle actually improves on the last across three deliberately different task types. The loop converged to a correct solution in one cycle on a solvable tool task, improved from 0.08 to 0.97 over five cycles on a task with exploitable feedback, and failed to improve at all on a task with no learnable signal. The finding is the contrast: autonomous self-improvement is real but bounded by whether the task exposes a gradient the loop can exploit. We report the plateau and the failure alongside the success.

1. Motivation

A common architecture diagram places 'continuous self-improvement' at the base of an autonomous agent, implying the system reliably gets better over time. In practice, no system improves unconditionally; improvement depends on the task providing feedback the agent can use. Rather than claim general self-improvement, we built the loop and measured its real behaviour, including where it stops improving. An evaluation that cannot fail cannot distinguish genuine autonomy from wishful description.

2. The loop

Each cycle runs: recall from memory; produce an action (invoke a matching tool, or compose from memory and the previous cycle's result); evaluate the action against the task's objective judge; reflect to estimate confidence and detect failure; and let a learning controller decide whether the attempt enters memory for future cycles. State carries forward, so a later cycle can build on what an earlier one established. The reasoning step runs behind the verification fence (TR-08), so no cycle proceeds on a fabricated intermediate claim.

3. Method

We ran the loop on three task types, each with an objective scorer returning a value in [0,1]: (A) a solvable arithmetic task a tool can compute exactly; (B) a refinement task with a hidden target and warmer/colder feedback the loop can exploit; (C) an adversarial task whose scorer is pure noise, offering no signal. Each task ran up to eight cycles.

4. Results

TaskScore trajectoryOutcome
A — solvable tool1.00converged, 1 cycle
B — refinement0.08 → 0.97improved, 5 cycles
C — adversarial~chancefailed to improve

On Task A the loop reached the correct answer immediately and stopped. On Task B it improved monotonically across cycles by feeding each result into the next attempt, climbing from near-zero to 0.97 — genuine autonomous refinement. On Task C, where the scorer carried no exploitable structure, the loop's scores fluctuated around chance and never improved; the loop correctly reported this as a plateau rather than manufacturing progress.

5. Interpretation

Self-improvement in the loop is entirely a function of exploitable feedback. Where the task exposes a gradient — a correct answer to reach, or a signal indicating warmer or colder — the loop climbs it. Where it does not, no amount of iteration helps, and the honest system says so. This bounds the 'continuous self-improvement' claim precisely: the mechanism is real and demonstrable, but it is not a general capability that improves on arbitrary tasks. It improves on tasks that are, in a specific sense, improvable.

6. Limitations

7. Conclusion

We built the autonomous loop the architecture calls for and measured its self-improvement honestly. It converges, it refines, and it plateaus — each on the appropriate task. The contribution is not a claim of open-ended self-improvement but a demonstration, with measurements, of exactly where autonomous iteration helps and where it cannot. For a system whose value proposition is trust, reporting the ceiling is the point.