How much agents may do on their own
Series: Aigile deep dives, paper 07 of 09. Elaborates working paper Section 13. Depends on Papers 03 and 06.
Abstract
Aigile spans the spectrum from human-AI pair collaboration to fully agentic delivery without forking into two methodologies. This paper specifies the mechanism: authority over verification is assigned per layer in a matrix, not per project in a mode switch; autonomy moves rightward through the matrix on evidence and moves back on incident, in the manner of graduated licensing; and one row never moves. It then develops the structural hazard specific to the right side of the matrix, correlated verification failure, and the independence requirements that mitigate it. The framing sentence for the whole paper: humans always own the why; the dial only ever moves who owns the check.
1. Position in the framework
The matrix operationalizes the core invariant. It supplies the deciding authority invoked by the deviation protocol (Paper 05, R16), the amendment authority of the constitution (Paper 03, R8), and the delegation trajectory along which the validation architecture’s human gates (Paper 06) remain fixed while everything around them moves.
2. The problem in full
2.1 The false binary
The intuitive framing is two setups: a collaborative mode, in which humans and agents both work the system actively and a competent human holds defining authority supported by AI input; and a fully agentic mode, in which agents hold the operational roles and humans contribute innovation input (feature intent) and lead architectural expertise (the constitution). Both descriptions are accurate as endpoints, and both are misleading as modes, because inspecting the “fully agentic” endpoint reveals that humans still own intent there too. What actually differs between the endpoints is not who owns the why but who performs each category of checking. That is not one switch. It is a setting per verification layer.
2.2 The trust question
Organizations adopting agentic delivery ask one question before all others: how do we trust this? A binary mode gives only dogmatic answers (trust it because the vendor does; distrust it because it is new). A per-layer matrix with evidential movement gives an operational answer: trust is granted where a track record exists, at the granularity where the record was earned, and revoked where an incident occurred. Trust becomes a paper trail instead of a vibe.
3. The mechanism
3.1 The matrix
| Verification layer | Collaborative end | Fully agentic end | Movement rule |
|---|---|---|---|
| Constitution amendments | Human decides; AI advises and drafts impact analysis | Lead-architect agent decides within a human-granted mandate | Partial: the mandate (appointment, scope, revocation) never leaves human hands |
| Feature intent and validation | Human | Human | Never moves (Paper 06, definitional) |
| Spec conformance | Human ritual, agent-assisted (Paper 05) | Standing conformance agent | Moves on the graduation trigger (evidence of outgrown ritual) |
| Behavioral verification (PR review) | Human reviews every atomic PR | Agent review with human sampling | Moves on track record; sampling never reaches zero |
| Mechanical verification (tests, CI) | Automated | Automated | Already right |
A project does not choose a mode. It has a matrix state, and “fully agentic” simply names the state in which most rows sit right. The matrix state is collaboration law and therefore lives in the constitution (Paper 03, content type five), which means changes to it are amendments: versioned, rationalized, and stamped.
3.2 Graduated autonomy
A verification responsibility moves rightward only after a defined track record, for example N consecutive increments in which human review found nothing the agent’s checks had missed. An incident moves the row left until the record is rebuilt. Grants and revocations are amendments, with rationale.
Three properties matter. Movement is per row, so a team can run agent PR review while keeping conformance ritual-based, matching trust to evidence at the layer where evidence exists. Movement is evidential in both directions, so autonomy is neither a reward for enthusiasm nor a permanent status; it is a licensed capability with a points system. And movement is recorded, so the answer to “why does the agent review PRs here” is a citable amendment with a track record attached, which is the audit answer regulated adopters need.
3.3 Correlated verification failure
Human peer review works partly on an unstated assumption: two humans have different brains, so their errors are weakly correlated, and the second pair of eyes adds real information. A builder agent and a reviewer agent sharing a base model share training, biases, and blind spots. Agent-on-agent review can therefore be confidently and consistently wrong: the reviewer approves the builder’s error because the reviewer would have made it too. The failure is worse than no review, because it launders error into confidence (the Paper 01 mechanism, relocated into the checking layer itself).
Mitigations, mandatory on the right side of the matrix and constitutional in force:
- Model diversity. Review and conformance roles run on a different model family than builder roles.
- Adversarial framing. Reviewers are prompted to find the ways this violates the specification, not to check whether it is fine; confirmation framing inherits the builder’s charity.
- Judgment-free checks first. Property-based tests, invariant checks, and executable criteria carry the load wherever they can, because they do not share anyone’s blind spots.
- Human sampling as tripwire. The sampling row never reaches zero; its purpose at high autonomy is not throughput but calibration, a standing estimate of what agent review misses, feeding the track-record accounting.
3.4 The fixed row
Feature intent and validation never move, for the definitional reason developed in Paper 06: the oracle for validation is human intent itself. This row is what makes the matrix one framework rather than a slippery slope; every other delegation is negotiable precisely because this one is not. In the fully agentic state, the humans’ remaining work is exactly the work the working paper assigns them: innovation in the form of feature intent, technical expertise in the form of the constitutional mandate, and validation with brains, eyes, and hands.
4. Normative protocol
R25 (Matrix residence). The authority matrix state is constitutional content; changes are amendments under R8. R26 (Evidential movement). Rightward movement requires a defined, recorded track record; incidents move the affected row left; both directions carry rationale. R27 (Fixed row). Feature intent and validation authority is human at every matrix state. Proposals to move it are inadmissible rather than merely rejected. R28 (Independence). At any state with agent verification, generation and verification are architecturally independent: distinct model families, adversarial framing, judgment-free checks preferred, and nonzero human sampling.
5. Orderly worked example
Orderly begins with all rows left. Month three: after 25 consecutive increments in which human PR review found nothing the agent’s checks had missed, an amendment (v2.4.0) moves behavioral verification to agent review with 20 percent human sampling; the reviewer agent runs on a different model family than the builders, adversarially framed. Month five: sampling catches an approved PR that weakened a retry criterion, an incident; the row moves back to full human review (v2.6.0), the track record restarts, and the retro finds the reviewer prompt had drifted toward confirmation framing, a Paper 08 collaboration learning. Month seven: the row re-graduates at 30 increments with corrected framing and sampling at 15 percent. Meanwhile the conformance row graduated in month four on the Paper 05 trigger, and the constitution row stays left throughout: Orderly’s humans keep amendment authority and have appointed no lead-architect agent, which is a legitimate permanent state, not a way station.
6. Failure modes
- Mode-switch thinking: adopting “fully agentic” as an identity and moving all rows at once. Counter: R26 makes undifferentiated movement impossible to justify.
- Monoculture review: builder and reviewer on the same model for convenience. Counter: R28 is constitutional, and monoculture discovered at audit is an incident.
- Track-record theater: counting increments without measuring what sampling missed. Counter: the record’s definition includes the sampling tripwire’s calibration data.
- Fixed-row erosion: validation quietly delegated via summaries and dashboards while nominally human. Counter: Paper 06’s brains-eyes-hands rule is the enforcement surface; a mediated validation is an unheld one.
7. Metrics
Matrix state over time (the maturity trajectory, per row); track-record lengths and incident counts per row; sampling catch rate (the calibration headline); model-diversity compliance; time-to-revocation after incident (should be near-immediate).
8. Open questions
Principled setting of N per row and per risk class; whether track records transfer across projects, teams, or agent generations (licensing suggests partially); mandate design for lead-architect agents (scope language, revocation triggers); insurance and liability framing of the matrix for enterprise adoption.
9. Derivative artifacts
Matrix template with movement-rule worksheets; amendment templates for grants and revocations; enterprise trust narrative for slides and website (the paper-trail answer to “how do we trust this”); audit-readiness mapping from R25 through R28 to common governance frameworks.