AI Systems · multi-agent workflow
FWEA — AI agent workflow
Builder → Reviewer → Fixer → Verifier, as I designed it. The review cycles shown here are read from its own log.
Claude Code · multi-agent design, review and verification
Multi-agent system designed and orchestrated by me.
Problem
An AI agent that grades its own work is not a reliable check: the same view that made a mistake is the one looking for it. FWEA separates building, reviewing, fixing and verifying into different agents, and records everything with hashes.
One agent alone
- builds the work
- grades its own work
- approves it
FWEA: separate roles
- a builder builds
- two isolated critics review
- a fixer repairs; a checker re-reads
- two fresh verifiers confirm
- the result is hashed
How I work
“I use AI agents as part of a structured development workflow: build → review → fix → verify, while keeping responsibility for requirements, decisions and final quality.”
ILF
Designed ILF, a personal framework of structured AI “labs” for web design, film, food and product work
Stated by Danilo
Architecture
The roles of the real FWEA run on 28 September 2026, as its review log records them. Each critic worked in isolation, and the tree that passed is pinned by its SHA-256 digest.
Builder → critics → fixer → verifiers
Evidence tier: Existing evidence · roles from the review log
The sequence, as text
- Builder builds the skill tree.
- Two critics, isolated and read-only, review it: lens 1 for correctness and decidability, lens 2 for scope and economy. Each reports BLOCKER, MAJOR and MINOR findings.
- Fixer corrects the findings; an internal checker re-reads the changes, and a fix-up closes what it finds.
- While a round fails, fresh critics review the corrected tree: rounds 0, 1 and 2, then the final critics.
- Two verifiers check that every serious finding of the final critics is closed, and look for regressions.
- PASS: the tree is pinned by its digest, sha256
8551e9a4a99b7b6b8e67c1530022adc2bfa15274545f57b50d1c842ab491593b.
Review cycles
Each critic reported its findings as BLOCKER, MAJOR or MINOR. 4 critic rounds ended FAIL; after the last fix round, two verifiers checked every serious finding and step 2 passed. The numbers are read from the log lines themselves when this site is built.
Findings per round and reviewer
Evidence tier: Existing evidence · parsed from the log
- BLOCKER
- MAJOR
- MINOR
- solid, hatched, hollow · numbers: BLOCKER · MAJOR · MINOR
| Round | Reviewer | BLOCKER | MAJOR | MINOR | Result | Log |
|---|---|---|---|---|---|---|
| Round 0 | Critic · Lens 1 | 3 | 12 | 11 | FAIL | line 17 of the review log, Round 0 |
| Critic · Lens 2 | 2 | 10 | 5 | FAILRound: FAIL | line 18 of the review log, Round 0 | |
| Round 1 | Critic · Lens 1 | 2 | 15 | 7 | FAIL | line 28 of the review log, Round 1 |
| Critic · Lens 2 | 0 | 4 | 8 | FAILRound: FAIL | line 29 of the review log, Round 1 | |
| Round 2 | Critic · Lens 1 | 1 | 11 | 10 | FAIL | line 39 of the review log, Round 2 |
| Critic · Lens 2 | 0 | 3 | 9 | FAILRound: FAIL | line 40 of the review log, Round 2 | |
| Final criticsstricter isolation | Critic · Lens 1 | 2 | 13 | 5 | FAIL | line 56 of the review log, Final critics |
| Critic · Lens 2 | 0 | 4 | 7 | FAILRound: 18 unique serious findings, all fixed | line 57 of the review log, Final critics | |
| Targeted verificationafter the last fix round | Verifier 1regression findings | 0 | 0 | 5 | CLEAN · 18/18 closed | line 67 of the review log, Targeted verification |
| Verifier 2regression findings | 0 | 0 | 4 | CLEAN · 19/19 closedRound: step 2 PASS | line 68 of the review log, Targeted verification |
Demo
A sample claim goes through the same four roles. It is a simulation: the sentence is invented and none of these findings come from the real log (the real numbers are in Review cycles).
A sample claim, step by step
Evidence tier: Simulation · illustrative
01 Builder: Writes a first draft
The builder turns the brief into a first version of the sentence.
Sample claim (draft) Expert in all modern web technologies, with flawless results.
Draft
02 Critics: Two isolated critics review it
Each critic reads only the draft and the brief, with read-only tools, and reports findings by severity.
Sample claim (draft) Expert in all modern web technologies, with flawless results.
- BLOCKER · Lens 1 · correctness & decidability: “all modern web technologies” cannot be decided: no evidence could ever prove it.
- MINOR · Lens 2 · scope & economy: “flawless results” is a vague phrase that adds nothing checkable.
FAIL · 1 BLOCKER, 0 MAJOR, 1 MINOR
03 Fixer: Fixes what the critics found
The fixer changes only what the findings name; an internal checker re-reads the changed lines.
Sample claim (after the fix) Expert in all modern web technologies, with flawless results.Builds responsive pages in HTML, CSS and JavaScript; each skill links to its evidence.- F1 (BLOCKER): the claim now names the skills, each with its evidence.
- F2 (MINOR): the vague phrase is removed.
- Checker: no new finding in the changed lines.
Fixed · waiting for verification
04 Verifier: A fresh verifier checks every finding
The verifier did not take part in the review or the fix. It checks each finding against the new text, then looks for regressions.
Sample claim (after the fix) Builds responsive pages in HTML, CSS and JavaScript; each skill links to its evidence.
- F1: CLOSED
- F2: CLOSED
- Regression: 0 BLOCKER, 0 MAJOR, 0 MINOR
PASS · sealed with its SHA-256
sha256 of the fixed sentence:
f0841dcaff317314bdd6e489d685ffa95469bac55e73e95c7411d50892da5a60
Evidence
The status of the skill, the checkpoint digest and manifest, and the verbatim lines of the review log that this page is built from.
Status
Evidence tier: Existing evidence
Release candidateBuilt and reviewed; final A/B validation not yet completed
Release candidate. Static review passed with two independent verifiers; the A/B benchmark against no-skill is designed but not run yet.
Multi-agent system designed and orchestrated by me.
Checkpoint
Evidence tier: Existing evidence
- Skill-tree digest (the tree that passed step 2)
8551e9a4a99b7b6b8e67c1530022adc2bfa15274545f57b50d1c842ab491593b- MANIFEST.sha256
- 684 lines · sha256
39f269514ec828c05051c5bbcb5b8b2536c5b2f4511c441250efd561c8bf96c7
A multi-agent AI workflow I designed and directed with Claude Code: it builds and tests a reusable web-design skill (builder, reviewer, fixer and verifier agents; checkpoints and SHA-256 hashes). The first version of this CV was built with it.
See proof
Claim
A multi-agent AI workflow I designed and directed with Claude Code: it builds and tests a reusable web-design skill (builder, reviewer, fixer and verifier agents; checkpoints and SHA-256 hashes). The first version of this CV was built with it.
Stated by Danilo
Evidence
- FWEA review log: the critics’ isolation check (tools/step2/STEP2_RUN_LOG.md, lines 10 and 12–14) · manual · checked 2 October 2026The isolation check at launch 1 (28/09/2026), read from the critic processes’ own start-up events: MCP servers, tools, permission mode. Copied verbatim on 02/10/2026 from the log (sha256 1beac5c6…); line 11 (the memory folder, a local path) is left out.
- FWEA review log: round 0 result (tools/step2/STEP2_RUN_LOG.md, lines 16–19) · manual · checked 2 October 2026Two isolated critics (lens 1 and lens 2) on the first tree, 28/09/2026: both FAIL. Copied verbatim on 02/10/2026 from the log (sha256 1beac5c6…).
- FWEA review log: round 1 result (tools/step2/STEP2_RUN_LOG.md, lines 27–30) · manual · checked 2 October 2026Fresh critics after correction round 1, 28/09/2026: both FAIL. Copied verbatim on 02/10/2026 from the log (sha256 1beac5c6…).
- FWEA review log: round 2 result (tools/step2/STEP2_RUN_LOG.md, lines 38–42) · manual · checked 2 October 2026Fresh critics after correction round 2, 28/09/2026: both FAIL; the trend of BLOCKER and MAJOR findings per lens. Copied verbatim on 02/10/2026 from the log (sha256 1beac5c6…).
- FWEA review log: final critics result (tools/step2/STEP2_RUN_LOG.md, lines 55–58) · manual · checked 2 October 2026Final critics with stricter isolation (no test material), 28/09/2026: both FAIL; 18 unique serious findings. Copied verbatim on 02/10/2026 from the log (sha256 1beac5c6…).
- FWEA review log: targeted verification result (tools/step2/STEP2_RUN_LOG.md, lines 66–70) · manual · checked 2 October 2026Two verifiers after the last fix round, 28/09/2026: every serious finding closed, no new BLOCKER or MAJOR; step 2 PASS. Copied verbatim on 02/10/2026 from the log (sha256 1beac5c6…).
- FWEA checkpoint record: skill-tree digest and MANIFEST.sha256 · manual · checked 2 October 2026Skill-tree digest 8551e9a4…593b (the tree that passed step 2) and MANIFEST.sha256 (684 lines, sha256 39f26951…96c7) of the release-candidate checkpoint: computed locally on 02/10/2026; the skill files are not published.
Result
Verbatim excerpts of the review log (28/09/2026): four rounds of two isolated critics each ended FAIL; after the last fix round, two verifiers closed every serious finding and step 2 passed. The skill is a release candidate: the A/B benchmark against no-skill is designed but not run yet. That Danilo designed and directed the workflow, and that the first version of this CV was built with it, are his statements.
AI-assisted development with Claude Code; multi-agent workflows with builder, reviewer, fixer and verifier roles; designing and testing Claude Code skills
See proof
Claim
AI-assisted development with Claude Code; multi-agent workflows with builder, reviewer, fixer and verifier roles; designing and testing Claude Code skills
Stated by Danilo
Evidence
- FWEA review log: the critics’ isolation check (tools/step2/STEP2_RUN_LOG.md, lines 10 and 12–14) · manual · checked 2 October 2026The isolation check at launch 1 (28/09/2026), read from the critic processes’ own start-up events: MCP servers, tools, permission mode. Copied verbatim on 02/10/2026 from the log (sha256 1beac5c6…); line 11 (the memory folder, a local path) is left out.
- FWEA review log: round 0 result (tools/step2/STEP2_RUN_LOG.md, lines 16–19) · manual · checked 2 October 2026Two isolated critics (lens 1 and lens 2) on the first tree, 28/09/2026: both FAIL. Copied verbatim on 02/10/2026 from the log (sha256 1beac5c6…).
- FWEA review log: targeted verification result (tools/step2/STEP2_RUN_LOG.md, lines 66–70) · manual · checked 2 October 2026Two verifiers after the last fix round, 28/09/2026: every serious finding closed, no new BLOCKER or MAJOR; step 2 PASS. Copied verbatim on 02/10/2026 from the log (sha256 1beac5c6…).
- FWEA checkpoint record: skill-tree digest and MANIFEST.sha256 · manual · checked 2 October 2026Skill-tree digest 8551e9a4…593b (the tree that passed step 2) and MANIFEST.sha256 (684 lines, sha256 39f26951…96c7) of the release-candidate checkpoint: computed locally on 02/10/2026; the skill files are not published.
Result
The FWEA review log (28/09/2026) and checkpoint record show a multi-agent workflow testing a skill: isolated critic agents with read-only tools, then two verifier agents that closed every serious finding. That Danilo designed the workflow, and the rest of this claim, are his statement.
The review log, verbatim
6 line ranges of the log, copied unchanged; the file itself is not published.
Evidence tier: Existing evidence
Isolation check at launch 1 (from the stream-json init events):
line 11 left out
- MCP servers: none;
- tools: Read/Grep/Glob plus session utilities; no Agent, Task, Skill, Bash, Edit, Write, WebFetch or WebSearch;
- permission mode: default.
Evidence tier: Existing evidence
## Round 0 result (launch 3, both critics valid: 0 void tool calls)
- L1 (082f4db933bb): FAIL — 3 BLOCKER, 12 MAJOR, 11 MINOR (17 tool calls, all inside packet)
- L2 (7807a7e16c55): FAIL — 2 BLOCKER, 10 MAJOR, 5 MINOR (18 tool calls, all inside packet; 279 s)
- Step 2 = FAIL on the pre-RC1 tree def5e846…. Correction round 1 of max 2 starts (Spec §13 Correction rounds).
Evidence tier: Existing evidence
## Round 1 result (fresh critics, both valid: 0 void tool calls)
- L1 (0db2ad0aa774): FAIL — 2 BLOCKER, 15 MAJOR, 7 MINOR (17 tool calls, all inside packet)
- L2 (12694d35367d): FAIL — 0 BLOCKER, 4 MAJOR, 8 MINOR (17 tool calls, all inside packet)
- Step 2 = FAIL on tree e2f346bc…. Correction round 2 of max 2 (last) starts; snapshot fwea-workspace/snapshots_pre_round2_skill. Fixer instructed with explicit anti-overfitting rules (critics can read the test spec).
Evidence tier: Existing evidence
## Round 2 result (fresh critics, both valid: 0 void tool calls)
- L1 (1bab4fd91aff): FAIL — 1 BLOCKER, 11 MAJOR, 10 MINOR (18 tool calls)
- L2 (17b649e28abc): FAIL — 0 BLOCKER, 3 MAJOR, 9 MINOR (19 tool calls)
- Trend L1 B/M: 3/12 → 2/15 → 1/11; L2 B/M: 2/10 → 0/4 → 0/3.
- Step 2 = FAIL after the 2 allowed correction rounds → OWNER DECISION required (Spec §13).
Evidence tier: Existing evidence
## Final critics (v1.1 isolation) — both valid, 0 void tool calls
- L1 (4d95e6389f06): FAIL — 2 BLOCKER, 13 MAJOR, 5 MINOR (14 tool calls)
- L2 (9bd3b623e9f7): FAIL — 0 BLOCKER, 4 MAJOR, 7 MINOR (15 tool calls)
- No finding argues from a test item (the critics had no test material). Residual serious findings: 2 BLOCKER + 17 MAJOR (L1-01 and L2-04 overlap, so 18 unique). They go to the owner one by one (D-FWEA-1); nothing is accepted automatically.
Evidence tier: Existing evidence
## Targeted verification result (D-FWEA-2) — STEP 2 PASS
- V1 (80df4d8e9a39): CLEAN — 18/18 findings CLOSED; regression 0 BLOCKER, 0 MAJOR, 5 MINOR; scripts sizes PASS, dead refs PASS, 0 duplicate groups (15 tool calls, 0 void).
- V2 (4fc38e62dc7a): CLEAN — 19/19 entries CLOSED (L2-04 counted separately); regression 0 BLOCKER, 0 MAJOR, 4 MINOR; scripts PASS (15 tool calls, 0 void).
- No finding among the 18 was accepted implicitly: all were fixed and independently verified closed.
- STEP 2 = PASS on tree 8551e9a4a99b7b6b8e67c1530022adc2bfa15274545f57b50d1c842ab491593b (2026-09-28).
Evidence tier: Existing evidence